WO2024046086A1 - 基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置 - Google Patents

基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置 Download PDF

Info

Publication number
WO2024046086A1
WO2024046086A1 PCT/CN2023/112465 CN2023112465W WO2024046086A1 WO 2024046086 A1 WO2024046086 A1 WO 2024046086A1 CN 2023112465 W CN2023112465 W CN 2023112465W WO 2024046086 A1 WO2024046086 A1 WO 2024046086A1
Authority
WO
WIPO (PCT)
Prior art keywords
oil
data
gas reservoir
production
model
Prior art date
Application number
PCT/CN2023/112465
Other languages
English (en)
French (fr)
Inventor
张吉群
贾德利
李欣
常军华
李夏宁
吴丽
王利明
崔丽宁
张洋
王全宾
Original Assignee
中国石油天然气股份有限公司
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 中国石油天然气股份有限公司 filed Critical 中国石油天然气股份有限公司
Publication of WO2024046086A1 publication Critical patent/WO2024046086A1/zh

Links

Classifications

    • 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
    • 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
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • 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
    • G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
    • 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

Definitions

  • the invention relates to the technical field of petroleum extraction, and specifically to an automatic history matching method based on RU-Net and LSTM neural network models, an automatic history matching device based on RU-Net and LSTM neural network models, and an electronic Device and a machine-readable storage medium.
  • the stochastic maximum likelihood estimation method is one of the sampling methods.
  • the sampling-based method is extremely time-consuming and is not suitable for large-scale problems. Multiple restart operations are required during the process. It is assumed that there is a linear relationship between the model parameters and the flow response during each update, and there is a certain error for nonlinear problems.
  • the set-based set smoothing algorithm no longer needs to restart the simulator and performs a one-time global update of the model parameters, thereby achieving the purpose of being faster than the Kalman filter.
  • the set smoothing algorithm since the set smoothing algorithm only updates the model parameters through one iteration, resulting in The degree of fitting is insufficient in the problem of reservoir history matching.
  • the uncertainty quantification method based on predictive focused analysis can be used to predict pollutant concentration data.
  • the standard functional component analysis method is used to reduce the dimensionality of the data and prediction variables, thereby solving the problems that arise in the prediction focus analysis. This method has been applied to oil and gas reservoir development and pollutant concentration prediction problems, but It is unable to effectively reduce the uncertainty of oilfield development indicator predictions, and cannot effectively guide oilfield development management and decision-making and carry out uncertainty quantification analysis.
  • the purpose of the embodiments of the present invention is to provide an automatic history matching method and device based on RU-Net and LSTM neural network models.
  • the automatic history matching method and device based on RU-Net and LSTM neural network models are used to solve the problem.
  • the above-mentioned manual history matching requires oil and gas reservoir engineers to have rich experience, and the process requires a lot of time to repeatedly adjust parameters, has a long research cycle, has limitations, and has high uncertainty in the prediction of oil and gas reservoir development indicators.
  • embodiments of the present invention provide an automatic history matching method based on RU-Net and LSTM neural network models, including:
  • Step 1 Grid the oil and gas reservoir attribute model by well partition: generate multiple well control units with the well control range of a single well layer as the boundary based on the well position coordinates in the oil and gas reservoir attribute model;
  • Step 2 Construct a data training set, including:
  • Step 21 For each well control unit:
  • Step 22 Perform parameter perturbation on the basic static data to generate multiple parallel calculation examples as input data for the data training set;
  • Step 23 Perform forward simulation on the multiple parallel examples generated through oil and gas reservoir numerical simulation to obtain the oil and gas reservoir pressure, saturation distribution and injection and production dynamics as the output data of the data training set;
  • Step 3 Process the sample data in the data training set
  • Step 4 Based on the processed sample data, train a prediction model for predicting oil and gas reservoir pressure, saturation distribution, and injection and production performance;
  • Step 5 Based on the prediction model, estimate the parameters of the oil and gas reservoir geological model through an intelligent optimization algorithm.
  • step 3 includes:
  • Step 31 Clean the sample data based on the production history of the production well, the production dynamics of surrounding wells, and the physical property distribution of the oil and gas reservoir;
  • Step 32 Eliminate outliers in the sample data
  • Step 33 Normalize the sample data.
  • step 33 includes:
  • Step 331 For the pressure field data in the sample data:
  • Step 332 For all data in the sample data except the pressure field, use the min-max normalization algorithm to perform normalization processing.
  • step 4 includes:
  • Step 41 Based on the processed sample data and the recursive fully convolutional neural network, train the pressure field prediction model and the saturation field prediction model;
  • Step 42 Based on the processed sample data and the long-short-term memory neural network, train the injection-mining production dynamic prediction model.
  • step 5 includes:
  • Step 51 Obtain the oil and gas reservoir geological model, oil and gas reservoir well control parameters and actual injection and production performance;
  • Step 52 Input the oil and gas reservoir geological model into the pressure field prediction model to obtain the pressure field of the oil and gas reservoir, and input the oil and gas reservoir geological model into the saturation field prediction model to obtain the saturation field of the oil and gas reservoir;
  • Step 53 Input the pressure field, the saturation field and the oil and gas reservoir well control parameters into the injection and production performance prediction model to obtain the predicted injection and production performance;
  • Step 54 Use the difference between the predicted injection and production production dynamics and the actual injection and production production dynamics as the loss function
  • Step 55 Use the stochastic maximum likelihood estimation algorithm and the grid adaptive direct search algorithm to perform stochastic optimization of parameters of the oil and gas reservoir geological model;
  • Step 56 Update the oil and gas reservoir geological model of the oil and gas reservoir area based on the corresponding parameters when the loss function is minimized.
  • Embodiments of the present invention also provide an automatic history matching device based on RU-Net and LSTM neural network models, including:
  • the grid segmentation module is used to grid segment the oil and gas reservoir attribute model by well partition: based on the well position coordinates in the oil and gas reservoir attribute model, multiple well control units are generated with the well control range of a single well layer as the boundary;
  • Data set building module used to build data training sets, including:
  • the data processing module is used to process the sample data in the data training set
  • the model building module is used to train prediction models for predicting pressure, saturation distribution and injection and production production dynamics based on processed sample data
  • the parameter estimation module is used to estimate the parameters of the oil and gas reservoir geological model through an intelligent optimization algorithm based on the prediction model.
  • the data processing module is specifically used for:
  • model building module is specifically used for:
  • the pressure field prediction model and the saturation field prediction model are trained;
  • the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • the processor executes the computer program, the above-mentioned steps are implemented.
  • Automatic history matching method based on RU-Net and LSTM neural network models.
  • the present invention provides a machine-readable storage medium. Instructions are stored on the machine-readable storage medium. The instructions are used to cause the machine to execute the above-mentioned automatic history matching method based on RU-Net and LSTM neural network models. .
  • This technical solution uses well partitions to perform grid segmentation, and constructs a data training set. Then, the sample data in the data training set is processed to obtain accurate sample data, and then based on the processed sample data, a model for predicting oil and gas reservoirs is trained. Prediction models of pressure, saturation distribution and injection and production performance, and finally based on the prediction model, parameter estimation of the oil and gas reservoir geological model is carried out through intelligent optimization algorithms.
  • the above method can overcome the shortcomings of artificial history matching methods such as reliance on engineer experience, long research cycle, and limitations. It can greatly shorten the research cycle, improve the prediction accuracy of dynamic data, and significantly reduce the uncertainty of oil and gas reservoir development index predictions. , achieve accurate description and prediction of the entire life cycle of oil and gas reservoir development, significantly improve the accuracy of characterization of reservoir heterogeneity, and reduce the uncertainty of remaining oil distribution.
  • Figure 1 is a flow chart of the automatic history matching method based on RU-Net and LSTM neural network models provided by the present invention
  • Figure 2 is a schematic diagram of the well control unit provided by the present invention.
  • Figure 3 is a schematic diagram of changes in water content indicators of different example wells provided by the present invention.
  • Figure 4 is a schematic diagram comparing the predicted injection and mining production dynamics and the actual injection and mining production dynamics provided by the present invention
  • Figure 5 is a schematic structural diagram of the automatic history matching device based on RU-Net and LSTM neural network models provided by the present invention.
  • Figure 1 is a flow chart of the automatic history matching method based on RU-Net and LSTM neural network models provided by the present invention
  • Figure 2 is a schematic diagram of the well control unit provided by the present invention
  • Figure 3 is a diagram of different calculation example wells provided by the present invention Schematic diagram of water content index changes
  • Figure 4 is the predicted injection and production production dynamics and actual injection provided by the present invention.
  • Figure 5 is a schematic structural diagram of the automatic history matching device based on RU-Net and LSTM neural network models provided by the present invention.
  • this embodiment of the present invention provides an automatic history matching method based on RU-Net and LSTM neural network models, including:
  • Step 1 Grid the oil and gas reservoir attribute model based on well partitions
  • step 1 based on the well location coordinates in the oil and gas reservoir attribute model, generate multiple well control units bounded by the well control range of a single well layer.
  • the oil and gas reservoir attribute model is constructed using conventional geological modeling steps, that is, first establishing a structural model of the oil and gas reservoir, and then establishing an oil and gas reservoir attribute model. It is easily available to those skilled in the art and will not be described in detail here. .
  • a complete oil and gas reservoir attribute model usually contains at least millions of grids
  • a complete oil and gas reservoir attribute model has a large number of attribute parameters. If the data of each grid is used, Calculation as input data will produce a disaster of dimensionality and a huge amount of data, resulting in too long data processing time and reduced data processing efficiency.
  • a single-layer well is taken as an example: with each well as the center, the oil and gas The reservoir attribute model is divided into grids, and the oil and gas reservoir attribute model is divided into multiple well control units. Each well control unit has a well, and each well control unit has multiple grids. Each well control unit The data is used as input data for calculation to reduce the amount of data processed.
  • Voronoi algorithm is used to determine the well control unit of each well in the geological model.
  • Voronoi algorithm is also called Thiessen polygon or Dirichlet diagram. It is composed of a set of continuous polygons composed of perpendicular bisectors of straight lines connecting two adjacent points. N points that are different on the plane divide the plane according to the nearest neighbor principle, and each point is associated with its nearest neighbor area.
  • a Delaunay triangle is a triangle formed by connecting related points that share an edge with an adjacent Voronoi polygon. The center of the circumcircle of a Delaunay triangle is a vertex of the Voronoi polygon associated with the triangle.
  • Step 2 Construct a data training set
  • Step 21 For each well control unit:
  • Step 22 Perform parameter perturbation on the basic static data to generate multiple parallel calculation examples as input data for the data training set;
  • Step 23 Perform forward simulation on multiple parallel examples generated through oil and gas reservoir numerical simulation to obtain oil and gas reservoir pressure, saturation distribution and injection and production dynamics as the output data of the data training set.
  • the static parameters include: permeability, porosity, thickness, initial oil saturation, initial pressure, etc. Therefore, each well control unit corresponds to permeability, porosity, thickness, initial oil saturation and initial pressure. Therefore, there are one million grids and two hundred wells in a certain oil and gas reservoir area, and it is a single layer wells, therefore, each grid corresponds to a type of parameter, so the oil and gas reservoir area has a total of one million permeability, one million porosity, one million thickness, one million initial If the above data is used to calculate the oil saturation and one million initial pressures, the amount of data will be huge. Therefore, parameter optimization is required to achieve the dimensionality reduction effect. Therefore, the center of each well in the two hundred wells is used to calculate the network.
  • the grid is divided into two hundred well control units, and the permeability, porosity, thickness, water body, fault, structure, initial oil saturation and initial pressure in each well control unit are averaged to obtain each static parameters corresponding to the well control units (a total of two hundred well control units), and then merge the data in the two hundred well control units as the basic static data of the oil and gas reservoir area, and finally reduce the dimension of one million data to two One hundred data, that is, there are only two hundred permeability, two hundred porosity, two hundred thickness, two hundred initial oil saturation and two hundred initial pressure in the basic static data of the entire oil and gas reservoir.
  • the area size of each well control unit is different, so the number of grids in each well control unit is also different.
  • Step 3 Process the sample data in the data training set
  • Step 31 Clean the sample data based on the production history of the production well, the production dynamics of surrounding wells, and the physical property distribution of the oil and gas reservoir;
  • Step 32 Eliminate outliers in the sample data
  • Step 33 Normalize the sample data.
  • Step 331 For the pressure field data in the sample data:
  • Step 332 For all data in the sample data except the pressure field, use the min-max normalization algorithm to perform normalization processing.
  • the sample data cleaning work is carried out. Based on the production history of the production well, the injection and production dynamics of surrounding wells, and the physical property distribution of the oil and gas reservoir, valid data are retained, invalid data are eliminated, and missing data are estimated. Convert dirty data into data that meets quality requirements; carry out sample data outlier detection, analyze the causes of outliers, and provide data anomaly warnings; carry out sample data normalization processing to prepare data for subsequent big data and machine learning algorithms Prepare.
  • the minimum-maximum normalization algorithm is directly used for normalization processing; for initial pressure data, each well control unit within each Calculate the average pressure value from the initial pressure of the time step, then average the pressure value in the pressure field data of each time step to obtain the pressure difference field, and then use the minimum-maximum normalization algorithm to calculate the pressure difference at each time step of the training set
  • the value field is normalized and the original pressure field data is detrended.
  • Step 4 Based on the processed sample data, train a prediction model for predicting oil and gas reservoir pressure, saturation distribution, and injection and production performance;
  • Step 41 Based on the processed sample data and the Recursive Fully Convolutional Neural Network (RU-Net), train the pressure field prediction model and the saturation field prediction model;
  • RU-Net Recursive Fully Convolutional Neural Network
  • Step 42 Based on the processed sample data and the long short-term memory neural network (LSTM), train the injection and mining production dynamic prediction model.
  • LSTM long short-term memory neural network
  • the sample data in the data training set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.
  • the regular term of the loss function of the pressure field prediction model and the saturation field prediction model uses the L2 norm
  • the regular term of the loss function of the injection and production dynamic prediction model uses the L1 norm.
  • the processed data training set can be divided into a first training data set for training the field prediction model and a second training data set for training the injection and production dynamic prediction model, and the first training data set is used as a recursive full prediction model respectively.
  • the input of the convolutional neural network is trained to obtain the pressure field prediction model and the saturation field prediction model.
  • the main parameters in the first training data set include; basic static data and a large number of parallel calculation examples, as well as the corresponding pressure field and sum obtained through forward simulation. Saturation field; use the second training data set as the input of the long short-term memory neural network to train the injection and production dynamic prediction model.
  • the second training data set mainly includes the corresponding pressure field, saturation field, Injection and production production dynamics and acquired historical well control parameters.
  • ADAM adaptive moment estimation optimization algorithm
  • SGD stochastic gradient descent
  • Step 5 Based on the prediction model, estimate the parameters of the oil and gas reservoir geological model through an intelligent optimization algorithm.
  • step 51 obtaining the oil and gas reservoir geological model, oil and gas reservoir well control parameters and actual injection and production performance;
  • Step 52 Input the oil and gas reservoir geological model into the pressure field prediction model to obtain the pressure field of the oil and gas reservoir, and input the oil and gas reservoir geological model into the saturation field prediction model to obtain the saturation field of the oil and gas reservoir;
  • Step 53 Input the pressure field, the saturation field and the oil and gas reservoir well control parameters into the injection and production performance prediction model to obtain the predicted injection and production performance;
  • Step 54 Use the difference between the predicted injection and production production dynamics and the actual injection and production production dynamics as the loss function
  • Step 55 Use the stochastic maximum likelihood estimation algorithm and the grid adaptive direct search algorithm to perform stochastic optimization of parameters of the oil and gas reservoir geological model;
  • Step 56 Update the oil and gas reservoir geological model of the oil and gas reservoir area based on the corresponding parameters when the loss function is minimized.
  • the well control parameters of oil and gas reservoirs include: water injection volume of water injection wells, water injection pressure, liquid production volume of production wells, etc.
  • the geological model of oil and gas reservoirs can be constructed based on previous well logging data, coring data, seismic data, etc., and the geological model contains a variety of static parameters that can reflect the physical properties of the oil and gas reservoir region, such as permeability, porosity and other data. That is, it corresponds to the static data in the data training set; the well control parameters can be set according to the actual oil production plan, etc.
  • geological modeling can be performed through well logging data, coring data, seismic data, etc., and It is conceivable to set well control parameters according to actual conditions. Therefore, the steps of how to establish a geological model and how to set well control parameters will not be described again here.
  • the parameters of the oil and gas reservoir geological model should be corrected to ensure the accuracy of the oil and gas reservoir geological model, so as to provide a basis for the subsequent prediction process. More specifically, when modifying the parameters of the oil and gas reservoir geological model, at least Modify one of the categories to achieve a more accurate geological model based on historical data fitting, which facilitates an accurate understanding of the underground structure of oil and gas reservoirs.
  • step 52 by inputting the geological model into the pressure field prediction model, the pressure field in the oil and gas reservoir area can be predicted; by inputting the geological model into the saturation field prediction model, the saturation field in the oil and gas reservoir area can be obtained.
  • Inputting the geological model into the field prediction model can also be understood as using the static parameters contained in the geological model, such as permeability, porosity and other data, as the input of the field prediction model to predict the pressure field and oil and gas reservoir area.
  • saturation field in step 53, use the pressure field predicted by the pressure field prediction model, the saturation field predicted by the saturation field prediction model, and the well control parameters as the trained injection and production dynamic prediction model. Input, the predicted injection and production production dynamics can be predicted.
  • the actual injection and production performance recorded during the actual oil production process is compared with the predicted injection and production performance. If the difference between the predicted injection and production performance and the actual injection and production performance is less than or equal to the preset threshold, it means that the two If the error is smaller, the predicted injection and production performance is more accurate, and there is no need to correct the geological model; if the difference between the predicted injection and production performance and the actual injection and production performance is greater than the preset threshold, it means that the errors between the two are relatively small. Large, the predicted injection and production performance is not accurate enough. The default is that there is a deviation between the input geological model (i.e. permeability, porosity and other parameters) and the actual permeability, porosity and other parameters in the oil and gas reservoir area.
  • the input geological model i.e. permeability, porosity and other parameters
  • the The difference between the predicted injection and mining production dynamics and the actual injection and mining production dynamics is used as the loss function, and the stochastic maximum likelihood estimation algorithm and the grid adaptive direct search algorithm are used to perform stochastic optimization of the static parameters of the geological model, and finally in the setting Within the number of iterations, use the static parameters corresponding to the minimization of the loss function in all iterations to update the oil and gas reservoir geological model, thereby obtaining the optimized oil and gas reservoir geological model.
  • the optimized oil and gas reservoir geological model includes permeability, pores Parameters such as degree are more in line with the current distribution characteristics of oil and gas reservoir areas. Therefore, they can provide a supporting basis for accurate prediction of subsequent injection and production dynamics.
  • maximum likelihood estimation provides a method to evaluate model parameters given observation data, that is: "the model is determined but the parameters are unknown".
  • the goal of the stochastic maximum likelihood algorithm (RML) is to find a set of model parameters that maximizes the posterior probability or minimizes the objective function.
  • the grid adaptive direct search algorithm (MADS) solves the minimization problem and inverts the model parameters. It is a type of derivative-free optimization algorithm.
  • the main idea of the algorithm is to divide each iteration into a search step and a detection step.
  • the search step is A global search is performed in the variable space and all feasible regions containing local optimum are identified; the detection step performs a local search in the variable space with the purpose of accurately finding the optimal solution.
  • the comparative analysis results of the predicted oil production and actual oil production in the past 350 months update the oil and gas reservoir geological model in the oil and gas reservoir area.
  • it is determined to conduct a comparative analysis of the predicted injection and production performance and the actual injection and production performance in July.
  • the geological model of the oil and gas reservoir is established to obtain the static parameters.
  • the geological model of the oil and gas reservoir is input into the field prediction model to predict the pressure field and saturation field of the oil and gas reservoir area.
  • the well control parameters and production of the water injection well in July are The well control parameters and the predicted pressure field and saturation field of the oil and gas reservoir area are used as the input of the injection and production performance prediction model to predict the injection and production performance in July, and then obtain the injection and production performance based on the same water injection well between July. Compare the actual injection and production performance obtained after oil production with the production well control parameters and the production well control parameters.
  • the difference between the predicted injection and production performance and the actual injection and production performance is less than or equal to the preset threshold, it means that the two The error is small, and there is no need to correct the geological model; if the difference between the predicted injection and production performance and the actual injection and production performance is greater than the preset threshold, it means that the error between the two is large, and the geological model needs to be corrected, that is,
  • the stochastic maximum likelihood estimation algorithm and the grid adaptive direct search algorithm are used to randomly modify the static parameters of the geological model and input them into the field prediction model to predict the pressure field in the oil and gas reservoir area and the saturation field in the oil and gas reservoir area.
  • the injection and production production performance is predicted, the records are saved, and the iterative cycle is carried out in sequence until After the number of iterations is reached, the predicted injection and production production dynamics are compared with the actual injection and production production dynamics, and the loss function is minimized, that is, the static parameters corresponding to the minimum difference between the predicted injection and production production dynamics and the actual injection and production production dynamics are selected.
  • Update the regional geological model of the oil and gas reservoir to obtain an updated geological model of the oil and gas reservoir, and conduct subsequent injection, production and production dynamic predictions based on the updated oil and gas reservoir geological model.
  • this embodiment of the present invention also provides an automatic history matching device based on RU-Net and LSTM neural network models, including:
  • the grid segmentation module 10 is used to perform grid segmentation of the oil and gas reservoir attribute model based on well partitions;
  • Data set construction module 20 used to build a data training set
  • the data processing module 30 is used to process the sample data in the data training set
  • the model building module 40 is used to train a prediction model for predicting pressure, saturation distribution and injection and production production dynamics based on the processed sample data;
  • the parameter estimation module 50 is used to estimate the parameters of the oil and gas reservoir geological model through an intelligent optimization algorithm based on the prediction model.
  • the present invention provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor.
  • the processor executes the computer program, the above-mentioned RU-Net-based method is implemented. and automatic history matching methods for LSTM neural network models.
  • the present invention provides a machine-readable storage medium. Instructions are stored on the machine-readable storage medium. The instructions are used to cause the machine to execute the above-mentioned automatic history matching method based on RU-Net and LSTM neural network models.
  • the program is stored in a storage medium and includes several instructions to cause the microcontroller, chip or processor to (processor) executes all or part of the steps of the methods described in various embodiments of this application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .
  • any combination of different implementation modes of the embodiments of the present invention can also be performed. As long as they do not violate the ideas of the embodiments of the present invention, they should also be regarded as the content disclosed in the embodiments of the present invention.

Landscapes

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

Abstract

本发明实施例提供一种基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置，属于石油开采技术领域。所述方法包括：以井分区进行油气藏属性模型的网格剖分；构建数据训练集；对数据训练集中的样本数据进行处理；基于处理后的样本数据，训练用于预测油气藏压力、饱和度分布和注采生产动态的预测模型；基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。本发明的基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置具有数据处理量小，大大缩短研究周期，能够精准预测油气藏的注采生产动态，显著提高储层非均质性刻画的精度，降低剩余油分布的不确定性，实现油气藏开发全生命周期的精确描述和预测的优点。

Description

基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置 技术领域
本发明涉及石油开采技术领域，具体地涉及一种基于RU-Net和LSTM神经网络模型的自动历史拟合方法、一种基于RU-Net和LSTM神经网络模型的自动历史拟合装置、一种电子设备及一种机器可读存储介质。
背景技术
油气藏数据历史拟合是油气藏模拟流程中极其重要的一环。通过人工进行历史拟合需要油气藏工程师具备丰富经验，且过程中需花费大量时间反复调整参数、研究周期长、具有局限性、油气藏开发指标预测的不确定性较高，因此，自动历史拟合方法成为近年来油气藏工程领域的研究热点。自动历史拟合方法可大致分为基于模型空间反演和基于数据空间反演两种思路。
在利用模型空间反演算法进行拟合时，随机最大似然估计方法是其中一种采样方法，但是，基于采样的方法，其过程极其耗时，所以不适用于大规模问题，且在拟合过程中需要进行多次重启动操作，假定模型参数和每次更新期间的流动响应之间存在线性关系，针对非线性问题时存在一定误差。基于集合的集合平滑算法不再需要重启模拟器，对模型参数进行一次性全局更新，从而达到了比卡尔曼滤波更快的目的，然而，由于集合平滑算法只通过一次迭代来更新模型参数，导致在油气藏历史拟合问题中的拟合程度不够。故发展了多重数据同化下的集合平滑算法，目前在油气藏历史拟合问题具有较好的应用效果。针对天然裂缝储层，开展历史拟合研究，但其采用的方法仅生成单个后验模型，从而不能有效降低油气藏指标预测的不确定性。目前，很难直接应用到实际裂缝储层的历史拟合问题中。
在利用数据空间反演算法时，现有技术中，可以基于预测聚焦分析的不确定性量化方法，用以预测污染物浓度数据。在此基础上，利用标准函数成分分析方法，将数据和预测变量进行降维处理，从而解决了预测聚焦分析中出现的问题，该方法已被应用于油气藏开发和污染物浓度预测问题，但是无法有效降低油田开发指标预测的不确定性，无法有效指导油田开发管理及决策和开展不确定性量化分析。
发明内容
本发明实施例的目的是提供一种基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置，该设备基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置用以解决上述的通过人工进行历史拟合需要油气藏工程师具备丰富经验，且过程中需花费大量时间反复调整参数、研究周期长、具有局限性、油气藏开发指标预测的不确定性较高的问题。
为了实现上述目的，本发明实施例提供一种基于RU-Net和LSTM神经网络模型的自动历史拟合方法，包括：
步骤1、以井分区进行油气藏属性模型的网格剖分：基于油气藏属性模型中的井位坐标生成多个以单井层的井控范围为边界的井控单元；
步骤2、构建数据训练集，包括：
步骤21、对于每一井控单元：
计算该井控单元内包含的所有网格中每一类静态参数的平均值，得到该井控单元的多类平均静态参数，将所有井控单元的多类平均静态参数作为基础静态数据；
步骤22、对基础静态数据进行参数扰动，生成多个平行算例，作为数据训练集的输入数据；
步骤23、通过油气藏数值模拟分别对生成的多个平行算例进行正演模拟，得到油气藏压力、饱和度分布和注采生产动态，作为数据训练集的输出数据；
步骤3、对数据训练集中的样本数据进行处理；
步骤4、基于处理后的样本数据，训练用于预测油气藏压力、饱和度分布和注采生产动态的预测模型；
步骤5、基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。
可选的，步骤3包括：
步骤31、基于生产井的生产历史、周边井的生产动态以及油气藏物性分布，进行样本数据清洗；
步骤32、剔除样本数据中的异常值；
步骤33、对样本数据进行归一化处理。
可选的，步骤33包括：
步骤331、对于样本数据中的压力场数据：
计算样本数据中所有时间步压力场的平均值；
将每一时间步的压力场数据中减去所有时间步压力场的平均值，得到压力差值场；
利用最小-最大归一化算法对所述压力差值场进行归一化处理；
步骤332、对于样本数据中除压力场的其余所有数据，利用最小-最大归一化算法进行归一化处理。
可选的，步骤4包括：
步骤41、基于处理后的样本数据和递归全卷积神经网络，训练得到压力场预测模型和饱和度场预测模型；
步骤42、基于处理后的样本数据和长短期记忆神经网络，训练得到注采生产动态预测模型。
可选的，步骤5包括：
步骤51、获取油气藏地质模型、油气藏井控参数和实际注采生产动态；
步骤52、将所述油气藏地质模型输入所述压力场预测模型，得到油气藏的压力场，将所述油气藏地质模型输入所述饱和度场预测模型，得到油气藏的饱和度场；
步骤53、将所述压力场、所述饱和度场和所述油气藏井控参数输入所述注采生产动态预测模型，得到预测的注采生产动态；
步骤54、将预测的注采生产动态和实际注采生产动态的差值作为损失函数；
步骤55、利用随机最大似然估计算法和网格自适应直接搜索算法，进行油气藏地质模型的参数随机优化；
步骤56、基于损失函数最小化时对应的参数更新油气藏区域的油气藏地质模型。
本发明实施例还提供一种基于RU-Net和LSTM神经网络模型的自动历史拟合装置，包括：
网格剖分模块，用于以井分区进行油气藏属性模型的网格剖分：基于油气藏属性模型中的井位坐标生成多个以单井层的井控范围为边界的井控单元；
数据集构建模块，用于构建数据训练集，包括：
对于每一井控单元：
计算该井控单元内包含的所有网格中每一类静态参数的平均值，得到该井控单元的多类平均静态参数，将所有井控单元的多类平均静态参数作为基础静态数据；
对基础静态数据进行参数扰动，生成多个平行算例，作为数据训练集的输入数据；
通过油气藏数值模拟分别对生成的多个平行算例进行正演模拟，得到油气藏压力、饱和度分布和注采生产动态，作为数据训练集的输出数据；
数据处理模块，用于对数据训练集中的样本数据进行处理；
模型建立模块，用于基于处理后的样本数据，训练用于预测压力、饱和度分布和注采生产动态的预测模型；
参数估计模块，用于基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。
可选的，所述数据处理模块具体用于：
基于生产井的生产历史、周边井的生产动态以及油气藏物性分布，进行样本数据清洗；
剔除样本数据中的异常值；
对样本数据进行归一化处理。
可选的，所述模型建立模块具体用于：
基于处理后的样本数据和递归全卷积神经网络，训练得到压力场预测模型和饱和度场预测模型；
基于处理后的样本数据和长短期记忆神经网络，训练得到注采生产动态预测模型。
另一方面，本发明提供一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现上述的基于RU-Net和LSTM神经网络模型的自动历史拟合方法。
另一方面，本发明提供一种机器可读存储介质，该机器可读存储介质上存储有指令，该指令用于使得机器执行上述的基于RU-Net和LSTM神经网络模型的自动历史拟合方法。
本技术方案以井分区进行网格剖分，并构建出数据训练集，再对数据训练集中的样本数据进行处理，得到准确的样本数据，然后基于处理后的样本数据训练出用于预测油气藏压力、饱和度分布和注采生产动态的预测模型，最后基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。通过上述的方法能够克服了人工历史拟合方法依赖工程师经验、研究周期长、存在局限性等缺陷，能够大大缩短研究周期，提升动态数据的预测精度，大幅降低油气藏开发指标预测的不确定性，实现油气藏开发全生命周期的精确描述和预测，显著提高储层非均质性刻画的精度，降低剩余油分布的不确定性。
本发明实施例的其它特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
附图是用来提供对本发明实施例的进一步理解，并且构成说明书的一部分，与下面的具体实施方式一起用于解释本发明实施例，但并不构成对本发明实施例的限制。在附图中：
图1是本发明提供的基于RU-Net和LSTM神经网络模型的自动历史拟合方法的流程图；
图2是本发明提供的井控单元的示意图；
图3是本发明提供的不同算例井含水指标变化示意图；
图4是本发明提供的预测的注采生产动态与实际注采生产动态的对比示意图；
图5是本发明提供的基于RU-Net和LSTM神经网络模型的自动历史拟合装置的结构示意图。
附图标记说明
10-网格剖分模块；             20-数据集构建模块；
30-数据处理模块；             40-模型建立模块；
50-参数估计模块。
具体实施方式
以下结合附图对本发明实施例的具体实施方式进行详细说明。应当理解的是，此处所描述的具体实施方式仅用于说明和解释本发明实施例，并不用于限制本发明实施例。
在本发明实施例中，术语“第一”、“第二”、“第三”等仅用于区分描述，而不能理解为指示或暗示相对重要性。
此外，“大致”、“基本”等用语旨在说明相关内容并不是要求绝对的精确，而是可以有一定的偏差。例如：“大致相等”并不仅仅表示绝对的相等，由于实际生产、操作过程中，难以做到绝对的“相等”，一般都存在一定的偏差。因此，除了绝对相等之外，“大致等于”还包括上述的存在一定偏差的情况。以此为例，其他情况下，除非有特别说明，“大致”、“基本”等用语均为与上述类似的含义。
图1是本发明提供的基于RU-Net和LSTM神经网络模型的自动历史拟合方法的流程图；图2是本发明提供的井控单元的示意图；图3是本发明提供的不同算例井含水指标变化示意图；图4是本发明提供的预测的注采生产动态与实际注 采生产动态的对比示意图；图5是本发明提供的基于RU-Net和LSTM神经网络模型的自动历史拟合装置的结构示意图。
实施例1
如图1所示，本发明实施例提供一种基于RU-Net和LSTM神经网络模型的自动历史拟合方法，包括：
步骤1、以井分区进行油气藏属性模型的网格剖分；
具体地，步骤1的具体步骤为：基于油气藏属性模型中的井位坐标生成多个以单井层的井控范围为边界的井控单元。所述油气藏属性模型采用常规地质建模的步骤构建而成，即：先建立油气藏的构造模型，再建立出油气藏属性模型，对本领域技术人员而言容易得到的，此处不再赘述。
另外，如图2所示，由于一个完整的油气藏属性模型通常包含有至少上百万个网格，因此，一个完整的油气藏属性模型具有大量的属性参数，如果以每一个网格的数据作为输入数据进行计算，会产生维数灾难，数据量巨大，造成数据处理时间过长，降低数据处理效率，在本实施方式中，以单层井为例：以每一个井为中心，对油气藏属性模型进行网格剖分，将油气藏属性模型划分为多个井控单元，每一个井控单元内具有一口井，每一井控单元内具有多个网格，以每一个井控单元的数据作为输入数据进行计算，以此来减少处理的数据量。
更具体地，在本实施方式中，以每一井为中心，利用Voronoi算法，在地质模型中确定每一井的井控单元。Voronoi算法又叫泰森多边形或Dirichlet图，它是由一组由连接两邻点直线的垂直平分线组成的连续多边形组成。N个在平面上有区别的点，按照最邻近原则划分平面，每个点与它的最近邻区域相关联。Delaunay三角形是由与相邻Voronoi多边形共享一条边的相关点连接而成的三角形。Delaunay三角形的外接圆圆心是与三角形相关的Voronoi多边形的一个顶点。
步骤2、构建数据训练集；
具体地包括：步骤21、对于每一井控单元：
计算该井控单元内包含的所有网格中每一类静态参数的平均值，得到该井控单元的多类平均静态参数，将所有井控单元的多类平均静态参数作为基础静态数据；
步骤22、对基础静态数据进行参数扰动，生成多个平行算例，作为数据训练集的输入数据；
步骤23、通过油气藏数值模拟分别对生成的多个平行算例进行正演模拟，得到油气藏压力、饱和度分布和注采生产动态，作为数据训练集的输出数据。
更具体地，通过步骤1进行网格剖分后，针对每一井控单元，计算该井控单元内的所有同一类型的静态参数的平均值，将该平均值作为该井控单元内该类型的静态参数，同理，对于其他类型的静态参数采用相同的处理办法，使得每一井控单元内每一类型的静态参数均只有一个，最后，将每一井控单元内每一类型的静态参数合并作为完整的基础静态数据，如图2所示，通过上述的处理方法能够大大减少数据的处理量，在后续训练和计算时，提高处理效率，减少处理时间。
在本实施方式中，静态参数包括：渗透率、孔隙度、厚度、初始含油饱和度和初始压力等。因此，每一井控单元对应的渗透率、孔隙度、厚度、初始含油饱和度和初始压力，因此，某一油气藏区域内设置有一百万个网格和两百口井，且为单层井，因此，每一个网格均对应有一个类型的参数，故该油气藏区域内一共具有一百万个渗透率、一百万个孔隙度、一百万个厚度、一百万个初始含油饱和度和一百万个初始压力，若以以上数据进行计算，数据量庞大，因此，需要进行参数优化，以达到降维效果，因此，以两百口井中每一口井的中心，对网格进行剖分，最终得到两百个井控单元，对每一个井控单元内的渗透率、孔隙度、厚度、水体、断层、构造、初始含油饱和度和初始压力求平均值，得到每一井控单元(共两百个井控单元)对应的静态参数，再将两百个井控单元中的数据合并作为油气藏区域的基础静态数据，最终便将一百万个数据降维度到两百个数据，即整个油气藏的基础静态数据中仅有两百个渗透率、两百个孔隙度、两百个厚度、两百个初始含油饱和度和两百个初始压力。另外，每一个井控单元的面积大小存在不同，因此每一井控单元内的网格数量也存在不同。
为了保证模型的准确度，需要利用大量的训练数据进行模型的训练。因此，在得到基础静态数据后，采用按照预设的倍数进行放大和缩小等方式，基于基础静态数据，对应的预设范围内进行参数扰动，得到大量平行算例，如图3所示，再通过油气藏数值模拟对大量平行算例数据进行正演模拟，能够得到不同输入数据组合下对应的模拟结果，即油气藏的压力场、饱和度场和注采生产动态，以此构建出数据训练集。
步骤3、对数据训练集中的样本数据进行处理；
具体包括：步骤31、基于生产井的生产历史、周边井的生产动态以及油气藏物性分布，进行样本数据清洗；
步骤32、剔除样本数据中的异常值；
步骤33、对样本数据进行归一化处理。
还包括：步骤331、对于样本数据中的压力场数据：
计算样本数据中所有时间步压力场的平均值；
将每一时间步的压力场数据中减去所有时间步压力场的平均值，得到压力差值场；
利用最小-最大归一化算法对所述压力差值场进行归一化处理；
步骤332、对于样本数据中除压力场的其余所有数据，利用最小-最大归一化算法进行归一化处理。
更具体地，在进行井控单元后，开展样本数据清洗工作，基于生产井的生产历史、周边井的注入及生产动态以及油气藏物性分布等，保留有效数据，剔除无效数据，估算缺失数据，将脏数据转化为满足质量要求的数据；展样本数据异常值检测工作，分析异常值出现原因，进行数据异常预警；开展样本数据归一化处理工作，为后续大数据与机器学习算法做好数据准备。除压力场数据外，对于渗透率、孔隙度、厚度、初始含油饱和度等，直接采用最小-最大归一化算法进行归一化处理；对于初始压力数据，将每一井控单元内每个时间步的初始压力求取平均压力值，再将每个时间步的压力场数据中平均压力值，得到压力差值场，然后再最小-最大归一化算法对训练集每个时间步压力差值场进行归一化，将原始压力场数据去趋势化。
步骤4、基于处理后的样本数据，训练用于预测油气藏压力、饱和度分布和注采生产动态的预测模型；
具体包括：步骤41、基于处理后的样本数据和递归全卷积神经网络(RU-Net)，训练得到压力场预测模型和饱和度场预测模型；
步骤42、基于处理后的样本数据和长短期记忆神经网络(LSTM)，训练得到注采生产动态预测模型。
更具体地，将数据训练集中的样本数据按照8:1:1的比例分为训练集、验证集和测试集。利用训练集训练模型参数，使用验证集调整神经网络超参数，依靠测试集评估模型效果。在训练过程，压力场预测模型和饱和度场预测模型的损失函数正则项使用L2范数，注采生产动态预测模型的损失函数正则项使用L1范数。
并且，处理后数据训练集可以分为用于训练场预测模型的第一训练数据集和用于训练注采生产动态预测模型的第二训练数据集，并将第一训练数据集分别作为递归全卷积神经网络的输入，训练得到压力场预测模型和饱和度场预测模型，第一训练数据集中主要参数包括；基础静态数据和大量平行算例，以及通过正演模拟得到的对应的压力场和饱和度场；将第二训练数据集作为长短期记忆神经网络的输入，训练得到注采生产动态预测模型，第二训练数据集主要包括通过正演模拟得到的对应的压力场、饱和度场、注采生产动态以及获取的历史井控参数。
在整个训练过程中，通过调整神经网络参数θ来最小化损失函数，损耗函数对于θ的梯度通过深度神经网络反向传播计算。优化时使用自适应矩估计优化算法(ADAM)，ADAM是一种可以替代传统随机梯度下降(SGD)过程的优化算法，基于训练数据迭代更新神经网络权重。循环神经网络具体的训练时间取决于许多因素，比如训练集大小、批量大小、优化器设置和学习速率，以及GPU性能。虽然训练时间可能因情况而异，但与油气藏历史拟合过程中数值模拟所需的时间相比，是非常短的。在训练过程中，学习速率、批量大小、训练迭代轮数和损失函数中定义的权重λ等几个重要的超参数需要进行反复调试。通过训练发现，两个网络均能在200轮数内收敛，饱和度通常比压力场训练收敛得快。通过数值实验发现，经过适当的数据预处理后训练过程受超参数值的设定影响较小。因此，该训练得到的超参数可用作新神经网络参数的初始设置。
步骤5、基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。
具体地包括：步骤51、获取油气藏地质模型、油气藏井控参数和实际注采生产动态；
步骤52、将所述油气藏地质模型输入所述压力场预测模型，得到油气藏的压力场，将所述油气藏地质模型输入所述饱和度场预测模型，得到油气藏的饱和度场；
步骤53、将所述压力场、所述饱和度场和所述油气藏井控参数输入所述注采生产动态预测模型，得到预测的注采生产动态；
步骤54、将预测的注采生产动态和实际注采生产动态的差值作为损失函数；
步骤55、利用随机最大似然估计算法和网格自适应直接搜索算法，进行油气藏地质模型的参数随机优化；
步骤56、基于损失函数最小化时对应的参数更新油气藏区域的油气藏地质模型。
更具体地，油气藏井控参数包括：注水井的注水量、注水压力、生产井的产液量等。油气藏地质模型可根据前期的测井数据、取心数据、地震数据等进行构建，并且地质模型中包含了多种能够反应油气藏区域物理属性的静态参数，如渗透率、孔隙度等数据，即对应数据训练集中的静态数据；井控参数可根据实际采油计划等情况进行井控参数设置，对本领域技术人员而言，通过测井数据、取心数据、地震数据等进行地质建模，以及根据实际情况设置井控参数是能够想到的，因此，如何建立地质模型和如何设置井控参数的步骤此处不再赘述。
但是，在实际运用过程中，由于在实际的注采生产动态预测过程中，油气藏地质模型在建立过程中有可能存在一定的误差，因此，可通过预测得到的注采生产动态和油气藏实际注采生产动态，对油气藏地质模型的参数进行修正，以保证油气藏地质模型的准确度，以便为后续的预测过程提供基础，更具体地，在进行油气藏地质模型的参数修正时，至少对其中一类进行修正，从而实现基于历史数据拟合得到更加精准的地质模型，便于准确的了解油气藏的地下构造情况。
因此，在步骤52中，将地质模型输入至压力场预测模型，能够预测得到油气藏区域的压力场；将地质模型输入至饱和度场预测模型，能够得到油气藏区域的饱和度场。将地质模型输入至场预测模型，也可以理解为是将地质模型中包含的静态参数，如渗透率、孔隙度等数据作为场预测模型的输入，预测得到油气藏区域的压力场和油气藏区域的饱和度场；在步骤53中，将通过压力场预测模型预测得到的压力场，通过饱和度场预测模型预测得到的饱和度场，以及井控参数作为训练好的注采生产动态预测模型的输入，能够预测得到预测的注采生产动态。通过实际采油过程中记录的实际注采生产动态与预测的注采生产动态进行对比判断，若预测得到的注采生产动态与实际注采生产动态之间的差值小于等于预设阈值，说明两者误差较小，预测的注采生产动态较为精准，无需对地质模型进行修正；若预测得到的注采生产动态与实际注采生产动态之间的差值大于预设阈值，说明两者误差较大，预测的注采生产动态不够精准，默认为输入的地质模型(即渗透率、孔隙度等参数)与实际的油气藏区域内的渗透率、孔隙度等参数之间存在偏差，因此，将预测得到的注采生产动态和实际注采生产动态的差值作为损失函数，并利用随机最大似然估计算法和网格自适应直接搜索算法，进行地质模型的静态参数随机优化，最终在设定的迭代次数内，利用所有迭代次数中损失函数最小化时对应的静态参数更新油气藏地质模型，从而得到优化后的油气藏地质模型，优化后的油气藏地质模型，其包含的渗透率、孔隙度等参数更加符合当前油气藏区域的分布特点，因此，能够为后续的注采生产动态精准预测，提供支撑基础。
另外，最大似然估计提供了一种给定观察数据来评估模型参数的方法，即：“模型已定，参数未知”。随机最大似然算法(RML)的目标是找到一组能够最大化后验概率或者最小化目标函数的模型参数。网格自适应直接搜索算法(MADS)求解最小化问题，反演模型参数，其是一类无导数优化算法，该算法的主要思想是将每次迭代分为搜索步和探测步，搜索步在变量空间中进行全局搜索并识别出所有包含局部最优的可行区域；探测步在变量空间中进行局部搜索，目的是精确查找最优解。
如图4所示，以采油量作为注采生产动态为例，以往350个月预测得到的采油量和实际采油量的对比分析结果更新油气藏区域的油气藏地质模型。在本实施方式中，以7月为例：确定对7月的预测的注采生产动态和实际注采生产动态进行对比分析，首先根据油气藏区域的测井数据、取心数据和地震数据，建立得到该油气藏地质模型，从而得到静态参数，将油气藏地质模型输入到场预测模型，预测得到该油气藏区域的压力场和饱和度场，并且，将7月的注水井井控参数和生产井井控参数以及预测得到的该油气藏区域的压力场和饱和度场作为注采生产动态预测模型的输入，预测出7月的注采生产动态，再获取7月之间基于相同注水井井控参数和生产井井控参数进行采油后得到的实际注采生产动态进行比对，若预测得到的注采生产动态与实际注采生产动态之间的差值小于等于预设阈值，说明两者误差较小，无需对地质模型进行修正；若预测得到的注采生产动态与实际注采生产动态之间的差值大于预设阈值，说明两者误差较大，需要对地质模型进行修正，即采用随机最大似然估计算法和网格自适应直接搜索算法，进行地质模型的静态参数随机修改，并输入至场预测模型，预测得到油气藏区域的压力场和油气藏区域内的饱和度场，再将新预测出压力场、饱和度场、注水井井控参数和生产井井控参数作为注采生产动态预测模型的输入，预测得到注采生产动态，进行保存记录，依次进行迭代循环，直至达到迭代次数后，进行预测注采生产动态与实际注采生产动态的对比，并筛选出损失函数最小化，即预测注采生产动态与实际注采生产动态之间差值最小时对应的静态参数更新油气藏区域地质模型，得到更新后的油气藏地质模型，并基于更新后的油气藏地质模型进行后续的注采生产动态预测。通过上述方案进行注采生产动态自动拟合，替代人工进行历史拟合，能够缩短研究周期、减小方法的局限性、降低油气藏开发指标预测的不确定性，显著提高储层非均质性刻画的精度，有效指导油田开发管理及决策和开展不确定性量化分析。
实施例2
如图5所示，本发明实施例还提供一种基于RU-Net和LSTM神经网络模型的自动历史拟合装置，包括：
网格剖分模块10，用于以井分区进行油气藏属性模型的网格剖分；
数据集构建模块20，用于构建数据训练集；
数据处理模块30，用于对数据训练集中的样本数据进行处理；
模型建立模块40，用于基于处理后的样本数据，训练用于预测压力、饱和度分布和注采生产动态的预测模型；
参数估计模块50，用于基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。
实施例3
本发明提供一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现上述的基于RU-Net和LSTM神经网络模型的自动历史拟合方法。
实施例4
本发明提供一种机器可读存储介质，该机器可读存储介质上存储有指令，该指令用于使得机器执行上述的基于RU-Net和LSTM神经网络模型的自动历史拟合方法。
以上结合附图详细描述了本发明实施例的可选实施方式，但是，本发明实施例并不限于上述实施方式中的具体细节，在本发明实施例的技术构思范围内，可以对本发明实施例的技术方案进行多种简单变型，这些简单变型均属于本发明实施例的保护范围。
另外需要说明的是，在上述具体实施方式中所描述的各个具体技术特征，在不矛盾的情况下，可以通过任何合适的方式进行组合。为了避免不必要的重复，本发明实施例对各种可能的组合方式不再另行说明。
本领域技术人员可以理解实现上述实施例方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成，该程序存储在一个存储介质中，包括若干指令用以使得单片机、芯片或处理器(processor)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
此外，本发明实施例的各种不同的实施方式之间也可以进行任意组合，只要其不违背本发明实施例的思想，其同样应当视为本发明实施例所公开的内容。

Claims (10)

  1. 一种基于RU-Net和LSTM神经网络模型的自动历史拟合方法，其特征在于，包括：
    步骤1、以井分区进行油气藏属性模型的网格剖分：基于油气藏属性模型中的井位坐标生成多个以单井层的井控范围为边界的井控单元；
    步骤2、构建数据训练集，包括：
    步骤21、对于每一井控单元：
    计算该井控单元内包含的所有网格中每一类静态参数的平均值，得到该井控单元的多类平均静态参数，将所有井控单元的多类平均静态参数作为基础静态数据；
    步骤22、对基础静态数据进行参数扰动，生成多个平行算例，作为数据训练集的输入数据；
    步骤23、通过油气藏数值模拟分别对生成的多个平行算例进行正演模拟，得到油气藏压力、饱和度分布和注采生产动态，作为数据训练集的输出数据；
    步骤3、对数据训练集中的样本数据进行处理；
    步骤4、基于处理后的样本数据，训练用于预测油气藏压力、饱和度分布和注采生产动态的预测模型；
    步骤5、基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。
  2. 根据权利要求1所述的方法，其特征在于，步骤3包括：
    步骤31、基于生产井的生产历史、周边井的生产动态以及油气藏物性分布，进行样本数据清洗；
    步骤32、剔除样本数据中的异常值；
    步骤33、对样本数据进行归一化处理。
  3. 根据权利要求2所述的方法，其特征在于，步骤33包括：
    步骤331、对于样本数据中的压力场数据：
    计算样本数据中所有时间步压力场的平均值；
    将每一时间步的压力场数据中减去所有时间步压力场的平均值，得到压力差值场；
    利用最小-最大归一化算法对所述压力差值场进行归一化处理；
    步骤332、对于样本数据中除压力场的其余所有数据，利用最小-最大归一化算法进行归一化处理。
  4. 根据权利要求1所述的方法，其特征在于，步骤4包括：
    步骤41、基于处理后的样本数据和递归全卷积神经网络，训练得到压力场预测模型和饱和度场预测模型；
    步骤42、基于处理后的样本数据和长短期记忆神经网络，训练得到注采生产动态预测模型。
  5. 根据权利要求4所述的方法，其特征在于，步骤5包括：
    步骤51、获取油气藏地质模型、油气藏井控参数和实际注采生产动态；
    步骤52、将所述油气藏地质模型输入所述压力场预测模型，得到油气藏的压力场，将所述油气藏地质模型输入所述饱和度场预测模型，得到油气藏的饱和度场；
    步骤53、将所述压力场、所述饱和度场和所述油气藏井控参数输入所述注采生产动态预测模型，得到预测的注采生产动态；
    步骤54、将预测的注采生产动态和实际注采生产动态的差值作为损失函数；
    步骤55、利用随机最大似然估计算法和网格自适应直接搜索算法，进行油气藏地质模型的参数随机优化；
    步骤56、基于损失函数最小化时对应的参数更新油气藏区域的油气藏地质模型。
  6. 一种基于RU-Net和LSTM神经网络模型的自动历史拟合装置，其特征在于，包括：
    网格剖分模块，用于以井分区进行油气藏属性模型的网格剖分：基于油气藏属性模型中的井位坐标生成多个以单井层的井控范围为边界的井控单元；
    数据集构建模块，用于构建数据训练集，包括：
    对于每一井控单元：
    计算该井控单元内包含的所有网格中每一类静态参数的平均值，得到该井控单元的多类平均静态参数，将所有井控单元的多类平均静态参数作为基础静态数据；
    对基础静态数据进行参数扰动，生成多个平行算例，作为数据训练集的输入数据；
    通过油气藏数值模拟分别对生成的多个平行算例进行正演模拟，得到油气藏压力、饱和度分布和注采生产动态，作为数据训练集的输出数据；
    数据处理模块，用于对数据训练集中的样本数据进行处理；
    模型建立模块，用于基于处理后的样本数据，训练用于预测压力、饱和度分布和注采生产动态的预测模型；
    参数估计模块，用于基于所述预测模型，通过智能优化算法进行油气藏地质模型的参数估计。
  7. 根据权利要求6所述的装置，其特征在于，所述数据处理模块具体用于：
    基于生产井的生产历史、周边井的生产动态以及油气藏物性分布，进行样本数据清洗；
    剔除样本数据中的异常值；
    对样本数据进行归一化处理。
  8. 根据权利要求6所述的装置，其特征在于，所述模型建立模块具体用于：
    基于处理后的样本数据和递归全卷积神经网络，训练得到压力场预测模型和饱和度场预测模型；
    基于处理后的样本数据和长短期记忆神经网络，训练得到注采生产动态预测模型。
  9. 一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现权利要求1-5中任一项所述的基于RU-Net和LSTM神经网络模型的自动历史拟合方法。
  10. 一种机器可读存储介质，该机器可读存储介质上存储有指令，该指令用于使得机器执行权利要求1-5中任一项所述的基于RU-Net和LSTM神经网络模型的自动历史拟合方法。
PCT/CN2023/112465 2022-08-31 2023-08-11 基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置 WO2024046086A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211055980.6A CN117669351A (zh) 2022-08-31 2022-08-31 基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置
CN202211055980.6 2022-08-31

Publications (1)

Publication Number Publication Date
WO2024046086A1 true WO2024046086A1 (zh) 2024-03-07

Family

ID=90081327

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/112465 WO2024046086A1 (zh) 2022-08-31 2023-08-11 基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置

Country Status (2)

Country Link
CN (1) CN117669351A (zh)
WO (1) WO2024046086A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117951627A (zh) * 2024-03-21 2024-04-30 潍柴动力股份有限公司 一种时序数据预测方法、装置及电子设备

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210224682A1 (en) * 2020-01-16 2021-07-22 Saudi Arabian Oil Company Training of machine learning algorithms for generating a reservoir digital twin
CN114357852A (zh) * 2021-11-12 2022-04-15 中法渤海地质服务有限公司 一种基于长短期记忆神经网络和粒子群优化算法的分层注水优化方法
CN114444402A (zh) * 2022-04-08 2022-05-06 中国石油大学(华东) 一种基于深度强化学习的油藏注采优化方法
CN114519274A (zh) * 2022-02-21 2022-05-20 中国石油大学(华东) 一种气驱油藏注采参数分步优化方法
CN114764546A (zh) * 2021-01-14 2022-07-19 中国石油化工股份有限公司 基于生成对抗神经网络的断块油藏剩余油预测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210224682A1 (en) * 2020-01-16 2021-07-22 Saudi Arabian Oil Company Training of machine learning algorithms for generating a reservoir digital twin
CN114764546A (zh) * 2021-01-14 2022-07-19 中国石油化工股份有限公司 基于生成对抗神经网络的断块油藏剩余油预测方法
CN114357852A (zh) * 2021-11-12 2022-04-15 中法渤海地质服务有限公司 一种基于长短期记忆神经网络和粒子群优化算法的分层注水优化方法
CN114519274A (zh) * 2022-02-21 2022-05-20 中国石油大学(华东) 一种气驱油藏注采参数分步优化方法
CN114444402A (zh) * 2022-04-08 2022-05-06 中国石油大学(华东) 一种基于深度强化学习的油藏注采优化方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
GU JIANWEI; REN YANLONG; WANG YIKE; LIU WEI: "Prediction Methods of Remaining Oil Plane Distribution Based on Machine Learning", ZHONGGUO SHIYOU DAXUE XUEBAO, vol. 44, no. 4, 20 August 2020 (2020-08-20), pages 39 - 46, XP009552948, ISSN: 1673-5005 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117951627A (zh) * 2024-03-21 2024-04-30 潍柴动力股份有限公司 一种时序数据预测方法、装置及电子设备

Also Published As

Publication number Publication date
CN117669351A (zh) 2024-03-08

Similar Documents

Publication Publication Date Title
CN109611087B (zh) 一种火山岩油藏储层参数智能预测方法及系统
WO2024046086A1 (zh) 基于RU-Net和LSTM神经网络模型的自动历史拟合方法及装置
CN111985610B (zh) 一种基于时序数据的抽油机井泵效预测系统和方法
CA2640725A1 (en) Methods, systems, and computer-readable media for fast updating of oil and gas field production models with physical and proxy simulators
CN108804382B (zh) 一种参数自动反求方法和装置
CN114896903A (zh) 一种基于强制学习的油田生产系统决策优化方法
KR101474874B1 (ko) 지능형 생산정 위치 선정 전산 시스템 및 이를 이용한 생산정 위치 선정 방법
CN110991079B (zh) 基于神经网络的油气藏参数解释方法、装置及电子设备
CN112113146A (zh) 供水管网管道粗糙系数和节点需水量同步自适应校核方法
CN113393051B (zh) 基于深度迁移学习的配电网投资决策方法
CN109670195B (zh) 融合单井敏感性局域化的EnKF油藏辅助历史拟合方法
Santos et al. An artificial intelligence method for improving upscaling in complex reservoirs
CN110096790B (zh) 一种基于强化学习的不规则缺陷漏磁信号反演方法
NO20200978A1 (en) Optimized methodology for automatic history matching of a petroleum reservoir model with ensemble kalman filter
Preis et al. Online hydraulic state prediction for water distribution systems
CN115564136A (zh) 一种地热历史拟合和产能预测方法
CN116838301A (zh) 一种开泵参数的获取方法及相关装置
CN112329232B (zh) 缝洞型油藏生产动态表征方法、装置、设备及存储介质
CN116906025B (zh) 一种采油工程生产异常诊断装置
CN114358441B (zh) 一种致密气产量的智能分段预测方法
CN116561942B (zh) 用于城市排水系统拓扑数据的修正方法以及装置
WO2024040801A1 (zh) 横波时差预测方法及装置
US10976468B2 (en) Methods of determining front propagation within a subsurface volume
CN117371983A (zh) 一种储罐大修周期预测方法、系统、存储介质和电子设备
CN116187178A (zh) 产量公式与机器学习双驱动的油井产量预测方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23859120

Country of ref document: EP

Kind code of ref document: A1