A sand production risk prediction method and system

Abstract

The present invention provides a sand production risk prediction method and system, which belongs to the field of oil and gas exploration and development technology and oil and gas field development engineering. The method comprises: the first step: collecting and cleaning the data of the target area; the second step: establishing a structured database to store the data obtained in the first step; the third step: using the data stored in the second step to establish multiple prediction models; the fourth step: optimizing each prediction model separately, and then synthesizing all the optimized prediction models into a final prediction model; the fifth step: using the final prediction model to obtain the sand production situation of the well to be predicted. The present invention can be used to quickly and efficiently establish a structured database for loose sandstone heavy oil and gas reservoirs, and establish a sand production risk prediction model based on different types of machine learning algorithms. The optimized prediction results are more accurate and more adaptable than prediction methods such as traditional experience or geostress models.

Description

Technical Field

[0001] The invention belongs to the field of oil and gas exploration and development technology and oil and gas field development engineering, and specifically relates to a sand production risk prediction method and system. Background Art

[0002] Effectively reducing and controlling the sand production in the wellbore is a long-standing research hotspot in the petroleum industry, and is also an important way to reduce costs and increase efficiency in oil and gas field production operations. Especially in loose sandstone heavy oil and gas reservoirs, the heavy oil will cause the sand and gravel in the loose rock mass to be stripped off due to its own high viscosity, resulting in sand production in the wellbore. This sand production phenomenon may lead to sand burial at the bottom of the well, increase the cost of well washing operations, reduce oil well production, erode production pipes and ground equipment, damage casing, and even scrap the oil well. Quantitative prediction of sand production is of great significance for optimizing sand control methods, production systems, increasing production and reducing development costs.

[0003] The sand production process of the formation is affected by reservoir physical properties, geomechanical characteristics, rock mechanics characteristics, completion methods, production systems, wellbore geometry characteristics, etc. Its response mechanism is complex under the influence of multiple factors. Traditional quantitative prediction methods rely on theoretical models such as rock strength failure criteria, which consider fewer factors and the prediction accuracy is difficult to meet production needs. There are many methods to determine the sand production of the wellbore. The more commonly used methods include the empirical method based on rock strength theory, which explains the sand production of oil wells as the change in rock strength caused by changes in pressure and production during the development of oil and gas; and the rock mechanics method of comprehensively analyzing whether shear or tensile failure occurs in the formation's in-situ stress and rock mechanical properties to determine whether the formation has sand.

[0004] However, both empirical methods and rock mechanics methods either consider a single factor, have great application limitations, and have low calculation accuracy, or have complicated modeling, long simulation cycles, and complex calculations. Therefore, it is necessary to explore a method that can make up for the shortcomings of both and realize the prediction and assessment method of sand production risk of loose sandstone heavy oil under the influence of high-dimensional variable nonlinearity.

[0005] Since 2007, the application of artificial intelligence (AI) has achieved technological breakthroughs in many industries such as medicine, finance, and transportation. In the exploration, development, and production business, the application of artificial intelligence is also growing rapidly. With the improvement of the information level of oil fields, various historical data have basically achieved structured storage, and a large amount of production information has been recorded. Machine learning technology has a strong ability to map high-dimensional and nonlinear problems. It is expected to use a large amount of historical information in oil fields to establish a quantitative prediction model for oil well sand production and guide the scientific optimization of production process measures. At home and abroad, the prediction of sand production in oil and gas reservoirs based on machine learning algorithms has been discussed and studied. At present, univariate nonlinear and multivariate linear regression methods based on collected geological, reservoir engineering and other data have been proposed. For example, the Chinese patent publication CN102278088A discloses a technology for solving sand production in loose sandstone heavy oil reservoirs. From the perspective of traditional data statistics, it proposes a correlation research method and prediction process for sand production in loose sandstone heavy oil reservoirs. This study first preprocessed and performed single-factor sensitivity tests on all horizontal well data in the study area, and then completed the feature screening of the main controlling factors, established and improved the structured database; at the same time, the decline curve analysis method (DCA) was used to calculate the production capacity prediction values ​​of all horizontal wells in the area. Then, mathematical methods such as regression methods were used to establish the correlation between sand production and geographical / reservoir physical / engineering parameters, and an optimal model for predicting the production capacity of heavy oil reservoirs was established. In order to improve the reliability of model training, the database training precision was expanded by cross-validation, and the prediction accuracy of the sand production model of heavy oil loose sandstone reservoirs was improved. However, few prediction methods based on machine learning technology have been proposed. The use of machine learning technology can not only predict long-term production capacity, but also take into account the optimization of short- and medium-term drilling and completion parameter design, which is very helpful for identifying the main control parameters affecting production in a specific time period and establishing and optimizing machine learning models. Summary of the invention

[0006] The purpose of the present invention is to solve the problems existing in the above-mentioned prior art, and to provide a sand production risk prediction method and system, which obtains the sand production amount in the oil well production process based on machine learning technology, and solves the problem of low prediction accuracy caused by the unclear mechanism of sand production in loose sandstone formations.

[0007] The present invention is achieved through the following technical solutions:

[0008] A first aspect of the present invention provides a method for predicting sand production risk, the method comprising:

[0009] Step 1: Collect and clean the data in the target area;

[0010] Step 2: Establish a structured database to store the data obtained in the first step;

[0011] Step 3: Use the data stored in step 2 to establish multiple prediction models;

[0012] Step 4: Optimize each prediction model separately, and then combine all optimized prediction models into the final prediction model;

[0013] Step 5: Use the final prediction model to obtain the sand production situation of the well to be predicted.

[0014] A further improvement of the present invention is that the operation of the first step includes:

[0015] Identify target areas;

[0016] Collect data from each well in the target area;

[0017] The collected data are supplemented and expanded.

[0018] The data of each well in the target area includes independent variables and dependent variables;

[0019] The independent variables include the following eight types of data:

[0020] The first type of data is well location geographic information, including: well location coordinates, reservoir depth, and wellbore dimensions;

[0021] The second type of data is well logging data, including: acoustic wave time difference, density, resistivity and natural potential;

[0022] The third type of data is reservoir characteristics, including: temperature, direction of minimum horizontal stress;

[0023] The fourth type of data is reservoir geomechanical characteristics, including: overburden pressure, pore pressure, maximum and minimum horizontal in-situ stress, and minimum horizontal in-situ stress orientation;

[0024] The fifth type of data is production parameters, including: daily oil production, daily water production, flow density, fluid viscosity, solution gas-oil ratio, bottom hole flow pressure and wellhead oil pressure;

[0025] The sixth category of data is completion parameters, including: perforation section length, perforation section depth, perforation density, phase angle, median particle size of gravel pack completion, seam width of wire-wound screen completion, seam width of metal wool sand filter and seam width of slotted screen;

[0026] The seventh type of data is steam injection parameters, including: cumulative gas injection volume, average steam injection temperature and average steam injection pressure;

[0027] The eighth type of data is wellbore trajectory, including: well inclination and azimuth information of horizontal wells;

[0028] The dependent variable is the sand production of a single well at a certain time point.

[0029] A further improvement of the present invention is that the operation of establishing a structured database in the second step includes:

[0030] A database with a two-dimensional table structure is used to store all the independent variables and dependent variables obtained in the first step;

[0031] The independent and dependent variables of each well are considered as a sample;

[0032] All samples are divided into two parts, one is training samples and the other is validation samples.

[0033] The multiple prediction models in the third step include:

[0034] Bayesian classification model, random forest model, support vector machine model and artificial neural network model.

[0035] The fourth step includes:

[0036] Firstly, each prediction model is optimized using training samples and validation samples to obtain each optimized prediction model;

[0037] Then, the model stacking method is used to iteratively process each optimized prediction model to obtain the final prediction model.

[0038] The operation of the fifth step includes:

[0039] Collect data of the well to be predicted, and process it according to the method in the first step to obtain the dependent variable of the well to be predicted;

[0040] The dependent variables of the well to be predicted are input into the final prediction model, and the final prediction model outputs the sand production situation of the well to be predicted.

[0041] A second aspect of the present invention provides a sand production risk prediction system, the system comprising:

[0042] A data acquisition and processing unit, used to collect and clean the data of the target area;

[0043] A database establishment unit, connected to the data acquisition and processing unit, is used to establish a structured database and store the data obtained by the data acquisition and processing unit;

[0044] A prediction model building unit, connected to the database building unit, for building a variety of prediction models using the data stored in the database;

[0045] A model synthesis unit, connected to the prediction model establishment unit, is used to optimize each prediction model established by the prediction model establishment unit, and then synthesize all the optimized prediction models into a final prediction model;

[0046] The prediction unit is connected to the data acquisition and processing unit and the model synthesis unit respectively, and obtains the sand production situation of the well to be predicted by using the data of the well to be predicted obtained by the data acquisition and processing unit and the final prediction model obtained by the model synthesis unit.

[0047] The database established by the database establishment unit is a database with a two-dimensional table structure.

[0048] The prediction models established by the prediction model establishment unit include: Bayesian classification model, random forest model, support vector machine model and artificial neural network model.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] By using the sand production risk prediction method based on multiple machine learning algorithms proposed in the present invention, a structured database for loose sandstone heavy oil and gas reservoirs can be established quickly and efficiently, and a sand production risk prediction model based on different types of machine learning algorithms can be established. The optimized prediction results are more accurate and more adaptable than traditional experience or geostress model prediction methods.

[0051] The machine learning model running in the Kaiyuan environment can quickly complete the training and prediction of sand production of all target samples (greater than or equal to 200 wells) in the area of ​​interest within 10 minutes, which can effectively assist in the selection of wellbore engineering parameters and production optimization of loose sandstone heavy oil reservoirs.

[0052] This method can calculate the influence weights of parameters such as geology, reservoir, lithology, and drilling and completion in loose sandstone heavy oil and gas layers, quickly identify the main controlling factors, and establish a high-dimensional machine learning prediction model. This model plays a key role in improving the quality and efficiency of oil and gas reservoir exploration and development, and helps loose sandstone heavy oil fields to use machine learning technology to accurately predict and evaluate the sand production risk in the region or in a single well. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flowchart of the steps of the method of the present invention;

[0054] Figure 2 Organize collection tables for oil field data for machine learning prediction;

[0055] Figure 3 This is a simple diagram that summarizes the logic of a single decision tree algorithm;

[0056] Figure 4 A simple diagram summarizing the logic of the random forest algorithm consisting of multiple decision trees;

[0057] Figure 5 A simple diagram that summarizes the logic of the artificial neural network algorithm;

[0058] Figure 6 A simple diagram to summarize the logic of combining machine learning models (Stacking);

[0059] Figure 7 A simple diagram to summarize the use of the "leave one out" method in cross-validation to optimize model accuracy;

[0060] Figure 8 It is a schematic diagram of the composition structure of the system of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0062] The present invention proposes a method for predicting the risk of sand production in the production process of loose sandstone heavy oil wells based on a machine learning algorithm. The method can calculate the influence weight values ​​of parameters such as geology, reservoir, lithology, and drilling and completion in loose sandstone heavy oil and gas layers, quickly identify the main controlling factors, and establish a high-dimensional machine learning prediction model. The model plays a key role in improving the quality and efficiency of oil and gas reservoir exploration and development, and helps loose sandstone heavy oil fields to accurately predict and evaluate the risk of sand production in the region or in a single well using machine learning technology. This method can improve the accuracy of sand control effect prediction, and the sand control effect can be improved by optimizing engineering design parameters.

[0063] From the perspective of data science, the present invention proposes a research method and a specific and feasible sand production risk prediction and assessment process for the correlation between the sand production risk of loose sandstone heavy oil layers and well location and wellbore information, reservoir physical properties, geomechanical characteristics, drilling and completion parameters, production parameters, and injection and production parameters. By collecting and cleaning the data of 594 heavy oil wells in the study area, a structured sand production risk assessment basic database is established, and the basic database is regionally divided into training and testing functions. Four machine learning algorithm models including Bayesian classification, random forest, support vector machine, and artificial neural network are established. After continuous improvement and verification through the test library to achieve the predetermined prediction accuracy, they are applied in the sand production prediction and assessment work of a single well.

[0064] like Figure 1 As shown, the method of the present invention comprises:

[0065] Step 1: Collect and clean the data in the target area;

[0066] Data collection is performed for the target loose sandstone heavy oil reservoir. The data can be tabular data from a public database, or it can be the operator's internal core laboratory report, drilling and completion daily report, and LAS file of logging and logging. The original database is entered by creating a compatible interface, automatic file recognition, or manual input. The present invention recommends using a tabular form. These entry methods are all mature technologies and will not be repeated here.

[0067] Specifically, the target area is first determined, and data collection and preprocessing are performed in the area. The data set is mainly divided into two parts: one is a set of parameters used as independent variables, including well location geographical characteristics, reservoir physical characteristics, geomechanical characteristics, steam extraction parameters, oilfield production parameters, completion engineering parameters and drilling engineering parameters; the other is the sand production of a single well at a certain time node, which is set to one year based on the experience of the target area. This type of parameter is called a dependent variable, also known as a label function.

[0068] Data types are mainly divided into major categories based on the different division of labor and functional attributes of traditional oil and gas field development. Figure 2 As shown in the figure, in the parameter set of independent variables: the first type of input parameters are well location geographic information, including well location coordinates, reservoir depth and borehole size. Take the well number or latitude and longitude coordinate calibration of a single well as an example: each training parameter or test parameter has and only has one geographic location coordinate. The influence degree is determined according to the absolute value of the predicted well location from the surrounding known training wells. Specifically, it is believed that the closer the training wells used for prediction are to the predicted wells, the greater the influence weight coefficient is, and vice versa; the reservoir depth is the vertical depth. Generally speaking, the greater the pressure, the greater the reservoir potential production energy, the greater the flow rate, and the greater the risk of sand production; the smaller the borehole size, the greater the corresponding pipeline flow rate, and the greater the risk of sand production. The second type of input parameters are logging data, including acoustic wave time difference, density, resistivity and natural potential data. The third type of input parameters are reservoir characteristics, mainly including parameters such as temperature and minimum horizontal stress direction. The fourth type of input parameters are reservoir geomechanical characteristics, mainly including: overburden pressure, pore pressure, maximum and minimum horizontal stress, and minimum horizontal stress orientation. The fifth type of input parameters are production parameters, mainly including: daily oil production, daily water production, flow density, fluid viscosity, solution gas-oil ratio, bottom hole flow pressure and wellhead oil pressure. The sixth type of input parameters are completion parameters, mainly including: casing perforation completion (perforation section length, perforation section depth, perforation density, phase angle), gravel packing completion (median particle size), wire wrap screen completion (slot width), metal wool sand filter (slot width) and slotted screen (slot width). The seventh type of input parameters are steam injection parameters, mainly including: cumulative gas injection volume, average steam injection temperature and average steam injection pressure. The eighth type of input parameters are wellbore trajectory, mainly including the well inclination and azimuth information of horizontal wells.

[0069] In summary, this method collects more than 30 independent variables related to well location geographical characteristics, reservoir physical properties, geomechanical characteristics, steam recovery, oilfield production, completion engineering and drilling engineering parameters.

[0070] Regarding data cleaning, the present invention mainly adopts two methods: data completion and data expansion.

[0071] The present invention attempts three missing data filling methods: mean / special value interpolation, K-nearest neighbor clustering and regression interpolation methods. After research, it is found that the mean or special value interpolation method is the easiest to implement and is also commonly used by people in the past, but it has great interference with the sample, especially when the interpolated value is regressed as an explanatory variable, the deviation between the estimated value of the parameter and the true value is large. In comparison, the clustering or regression interpolation method has a good effect on the interpolation of the missing value type as random missing. Therefore, the present invention focuses on using the latter two methods for data filling. Such as hot card filling, filling using all possible values, combined completeness method and maximum likelihood estimation and multiple interpolation methods.

[0072] For the sand production prediction, the present invention attempts two methods to expand missing data: using classical empirical formulas to expand data and using the information in this article to interpret and expand data. Specifically, with the help of formulas, such as the decline formula of a single well and the short-term production of a single well, the production status of a single well in the next 1 to 3 years or even longer can be derived, thereby artificially expanding the production subtable data; on the other hand, a large amount of text information data generated in the development and production of oil and gas fields can also be used. One-hot encoding can be used to deeply mine text information and digitize its features, thereby supplementing the database capacity.

[0073] Step 2: Establish a structured database to store the data obtained in the first step: store all independent variables and dependent variables in a two-dimensional table database, and divide the samples into two parts, one for training the model and the other for testing the model, as follows:

[0074] According to the results of the analysis of sand production influencing factors, the data design for sand production prediction was completed. According to the data design, an Access database for sand production prediction was established to comprehensively manage the full set of geological and engineering data of 250 wells in 5 oil fields, providing data for the research of sand production prediction methods based on machine learning. It is worth noting that there are 250 research samples in the database, and each sample contains two important variable sets: the first is the independent variable set, which mainly includes the 8 major types of feature information collected in the first step; the second is the sand production situation of each sample, that is, sand production or no sand production, which will be used as the dependent variable for subsequent machine learning training, testing and prediction.

[0075] There are many factors that lead to sand production in loose sandstone heavy oil layers, so it is relatively complicated to accurately predict it. Many models have been constructed for sand production prediction: For example, the empirical method, based on rock strength theory, explains the sand production in oil wells as the change in rock strength caused by changes in pressure and production during the development of oil and gas. When the rock strength is lower than the force exerted on the rock, the sandstone will break and cause the movement of sand. Sand production will cause sand arches in the wellbore. The sand arch has a certain bearing capacity. When the sand arch is destroyed, a large amount of formation sand will flow into the wellbore and be brought to the surface by the fluid to form sand. The rock mechanics method is to judge whether the formation has sand production by comprehensively analyzing whether shear or tensile failure occurs in the formation stress and rock mechanical properties; the machine learning method has a strong ability to describe nonlinear problems, which improves the accuracy of sand production prediction in oil wells.

[0076] Whether it is empirical method, rock mechanics method or machine learning method, it is necessary to build an institutionalized database that can be oriented to oil and gas fields. This database should have the function of real-time data transmission and a certain ability to update and correct data. After screening and verification (for example: data with a transmission value of 0, empty or unclear will be uniformly marked as NA value; for example: data with extreme abnormal values ​​of 99999999 will also be processed as NA value; for example, if the NA value of the independent variable in a sample exceeds 30%, it will be directly deleted, etc., all of which are implemented using existing technologies and will not be repeated here.) the independent variables and dependent variables will be used as input parameters to prepare for the establishment of the next machine learning model.

[0077] The database is a resource library of oil and gas field data. After determining the prediction target, relevant data materials are selected from the database, and then a mathematical model of classification, clustering or regression is established using a machine learning algorithm to achieve the target. Specifically, the database established by the present invention includes a central master table and multiple sub-tables. The main information contained in the master table is: the identity number of the sample well (each well has only one, and they are not repeated), the well name, the geographic coordinates, the oil field or oil region location and other descriptive information. The descriptive information of all sample wells is in the master table; each sample well corresponds to multiple sub-tables, and the sub-tables are divided into completion information table, well trajectory table, reservoir attribute table, geological information table, drilling engineering parameter table, fracturing parameter table, logging data table, etc. according to different disciplines. The specific information included in each sub-table can be designed according to actual needs.

[0078] The database of the present invention is a database that is logically expressed and implemented through a "two-dimensional table structure" and strictly follows the data format and length specifications. The parent table and the sub-table are unified to form a large database, and the required table combination is selected from it for different research objectives. For example, to predict the simple single-well production decline rate, only the parent table plus the single-well production sub-table is selected; for example, to predict the sand production risk, what needs to be selected is the parent table plus multiple sub-tables such as production system, single-well production, and rock mechanics parameters.

[0079] In a structured database, all parameters used to predict sand production risk are called "independent variables". Assume that the database contains data m∈M horizontal well set, n∈N independent variable set, where m is a single well sample, such as m1 is well 1, m2 is well 2, and so on until mi is the i-th well. The collection of all m is M, that is, the horizontal well collection; and n is the description of each independent variable, such as n1 is porosity, n2 is rock stress value, etc., and the collection of all independent variables is N. In this way, it is possible to predict the risk from x i ∈R n Sand production at a specific time point (for a certain oil field in the sub-database) Among them, xi represents the key independent variable selected from N, R n Represents all independent variables in N, y represents the objective function, which can be production, capacity or sand production. R is the total set of sub-databases of the ith well, because in different application scenarios, all data of the ith well are not needed, but only the data that are of interest or related to the research target. The set of these data is part of the database, so it is called a "sub-database".

[0080] This method belongs to "supervised regression learning", where y i is called x i The set of independent variables and corresponding labels is denoted as P = {P1, ..., P m}, where P i =(x i ,y i )=(x i1 , ..., x in ,y i )∈R n+1 (represents all sets of independent variables for each well in the sub-database). In addition, the independent variable sets and corresponding label sets that are classified separately are also described as X={X1,...,X m} (representing all independent variables) and Y = {Y1, ..., Y m} (represents all labels). Note that reusing data in P, X, and Y is allowed when using multi-set mathematical expressions.

[0081] In order to train and validate the model, the data in the P set needs to be properly partitioned. To achieve this goal, t sub-libraries / cardinality are designed, 1<t<m, and all selected parameters are derived from Set (T is a subset of M, representing the training set). T The number set is a training number set and belongs to the P set, P V is called the set of independent variables, belonging to P and related to P T is called complementary (i.e. P T is the training sample, P V is the validation sample, and the two together are the entire sample), P V =P\P T ={pv1, ..., pv k}. Finally, by comparing and predicting y′ i (that is, input the verification sample into the model, and the predicted value output by the model) and the label y i The difference between the value (i.e. the label corresponding to the validation sample itself) is called the loss or error function to judge the quality of the model. Well-known loss functions include standard deviation (Root Mean Squared Error: RMSE), mean squared error (Mean Squared Error: MSE) and Spearman correlation coefficient (Pearson and Spearman Correlation Coefficient).

[0082] Step 3: Application of machine learning algorithms: Build four prediction models, as follows:

[0083] The core of machine learning is to find the relationship between the independent variables and the dependent variables of the data. The main reason for its development is the development of a large amount of data. The traditional data analysis method can no longer process so much seemingly unrelated data. Therefore, data mining technology is needed to extract the relationship between various data and variables, so as to refine the data.

[0084] Machine learning is a method that can give computers the ability to learn so that they can complete functions that cannot be completed by direct programming. In a practical sense, machine learning is a method that uses data to train models and then uses models for prediction. Conventional machine learning methods can obtain the intrinsic connections in data through self-training without the need for artificial input of prior knowledge, thus avoiding the misleading effect of artificial input of prior knowledge (such as theoretical or empirical formulas) on data mining. Machine learning can be used for tasks such as classification, regression, clustering and dimensionality reduction. Mechanical drilling rate prediction is a typical regression problem. Commonly used machine learning classification algorithms include k-nearest neighbor method, support vector machine, random forest, BP neural network, etc. On the basis of a structured database, the present invention respectively analyzes the correlation between input parameters and label functions by establishing four machine learning models: Bayesian classification, random forest classification, support vector machine and artificial neural network, so as to predict the sand production risk size at a specific time node.

[0085] Among them, Bayesian classification is a statistical classification method based on Bayesian theorem. It classifies by predicting the probability that a given tuple belongs to a specific class.

[0086] 1) Naive Bayes classification model

[0087] Bayesian classification is a statistical classification method based on Bayes' theorem. It classifies by predicting the probability that a given tuple belongs to a specific class. Naive Bayesian classification assumes that the influence of an attribute value on a given class is independent of other attributes - class conditional independence. Its advantages are that it requires fewer parameters to be estimated and is insensitive to missing data, but it still contains related disadvantages: it assumes that attributes are independent of each other and requires knowing prior probabilities.

[0088] 2) Support Vector Machine (SVM) Model

[0089] The support vector machine adopts the principle of structural risk minimization, transforms the input feature space into a high-dimensional space through the nonlinear transformation of the kernel function, and finds the optimal classification hyperplane in the high-dimensional space. It is more suitable for small sample, nonlinear and high-dimensional classification problems. Since the computational complexity of SVM with kernel function is relatively large under large data volume, the present invention selects linear support vector machine as the evaluation method for the classification of the invention.

[0090] 3) Random Forest (RF) Model

[0091] Random forest (such as Figure 4 ) is a decision tree based on Figure 3The basic idea of ​​the ensemble method is to combine multiple weak classifiers into a strong classifier. This method has the characteristics of high prediction accuracy, strong generalization ability, low sensitivity to outliers and noise in the regression and prediction of industrial data, and has fewer hyperparameters and simple parameter adjustment. It is widely used in various industrial scenarios.

[0092] 4) BP Neural Network (Back Propagation) Model

[0093] BP neural network (such as Figure 5 The concept of artificial neural network (as shown in Figure 1) was proposed by scientists led by Rumelhart and McClelland in 1986. It is a multi-layer feedforward neural network trained according to the error back propagation algorithm and is currently the most widely used neural network. In the fully connected structure, the lower layer neurons are connected to all the upper layer neurons, which is the most common artificial neural network structure. The multi-layer network structure helps to extract and characterize the input features in layers, but the full connection of neurons will lead to too large a number of parameters and an increase in the amount of training.

[0094] In a fully connected structure, the lower layer neurons are connected to all the upper layer neurons, which is the most common artificial neural network structure. Multi-layer network structure helps to extract and characterize input features in layers, but fully connected neurons will lead to too large a number of parameters and an increase in training workload.

[0095] Figure 4 and Figure 5 The structure of the above two machine learning models is explained by taking capacity prediction as an example.

[0096] The above four machine learning prediction models are all existing mature technologies and will not be elaborated here.

[0097] Step 4: Optimize the model to obtain the final prediction model: First optimize each prediction model separately, and then synthesize the four models into the final prediction model, as follows:

[0098] All the sand prediction models used for optimization, whether Bayesian classification, random forest, or artificial neural network methods, are based on the same input parameters and label functions, and the data ratios of the training library and the test library are also consistent, that is, the same parameters are input to the four models, and the training samples and test samples are exactly the same.

[0099] In addition, in order to improve the reliability of model training, the present invention adopts the "leave one out" method in the cross-validation method (such as Figure 7The optimized sand prediction method and evaluation strategy can be applied to loose sandstone. The test case shows that the accuracy of production capacity classification prediction using machine learning methods reaches more than 85%. This study provides a strong guarantee for accurate evaluation and rapid business decision-making of loose sandstone heavy oil reservoirs from the perspective of data mining and machine learning.

[0100] Based on the four machine learning algorithms mentioned in the third step, namely, Bayesian classification, random forest classification, support vector machine and artificial neural network model, the Stacking method (model stacking method) is adopted (such as Figure 6 As shown in Figure 2, a composite machine learning model is built using a variety of methods to improve the model.

[0101] The details are as follows:

[0102] In the first stage, the conventional model building process was used to complete and optimize the establishment of four different machine learning prediction models. The goal of this stage was to optimize a single model.

[0103] The second stage is to iteratively process the four individually optimized models using the model stacking method to obtain a composite black box model. The goal of this stage is similar to the "leave one out" method, which can not only give full play to the accuracy prediction advantages of the superposition of the four models, but also improve the stability of model prediction. However, if the amount of data is too small, the training set selected from the P set will be deeply affected by the loss function during training and testing, and the prediction results will be inaccurate. In order to obtain relatively reliable error statistics in small or medium-sized databases, such as the present invention, cross-validation techniques will be used simultaneously to make up for the lack of data, such as Figure 7 As shown in Figure 1, this method collects data from the P set multiple times and uses different data combinations to create different training sets and test sets. Based on each dependent variable set, the loss function of the predicted results is averaged, and the average value is called the model f T ; λ is the generalization error. The final learning algorithm is optimized based on adjusting relevant parameters in order to improve the prediction accuracy.

[0104] like Figure 7 As shown in the figure, the upper part of the entire model is a basic machine learning model and a 5-fold cross validation is performed. For example, using Naive Bayes as the basic model 1, the 5-fold cross validation is to first take out four folds as the training set and the other fold as the test set, that is, Figure 7 Subsets 1 to 5 in ( Figure 7The 5 categories in represent 5 different test set & validation set classifications). Note: In stacking, this part of the data will use the entire training set. For example, assuming that the entire training set contains 10,000 rows of data and the test set contains 2,500 rows of data, then each cross-validation is actually a division of the training set. In each cross-validation, the training data will be 8,000 rows and the test data will be 2,000 rows. Each cross-validation includes two processes: 1. Training the model based on the training samples; 2. Predicting the test samples based on the model generated by the training sample training. After the first cross-validation is completed, the predicted value of the current test sample will be obtained. This will be a one-dimensional data of 2,000 rows, denoted as a1. After this part of the operation is completed, the entire original test sample of the sample set must be predicted. This process will generate 2,500 predicted values, which will be used as part of the test sample of the next layer of model, denoted as b1. Because a 5-fold cross validation is performed, the above-mentioned process will be performed five times, and finally 5 columns and 2000 rows of data a1, a2, a3, a4, a5 will be generated for the test set sample prediction, and the prediction for the test sample will be 5 columns and 2500 rows of data b1, b2, b3, b4, b5. After completing the entire step of Model1, it can be found that a1, a2, a3, a4, a5 are actually the predicted values ​​of the original entire training set. Putting them together will form a matrix with 10,000 rows and one column, recorded as A1. For the data of b1, b2, b3, b4, b5, the average value of each part is added to obtain a matrix with 2500 rows and one column, recorded as B1. The above cross-validation method is a mature technology and will not be repeated here.

[0105] The stacking of the present invention also includes multiple other models, including Model 2: support vector machine, Model 3: random forest, and Model 4: BP neural network. For these three models, the above steps can be repeated, and after the entire process is completed, new A2, A3, A4, B2, B3, and B4 matrices can be obtained. Similarly, all different machine learning models are optimized separately and then stacked into a composite black box model using the stacking method, which is the final prediction model.

[0106] Step 5: Use the final prediction model to predict the sand production of the well to be predicted:

[0107] Collect data of the well to be predicted, and process it according to the method in the first step to obtain the dependent variable of the well to be predicted;

[0108] The dependent variables are input into the final prediction model, and the prediction model outputs the sand production situation of the well to be predicted.

[0109] like Figure 8As shown, the present invention also provides a sand production risk prediction system, the system comprising:

[0110] A data acquisition and processing unit 10 is used to collect and clean the data of the target area;

[0111] A database establishment unit 20, connected to the data acquisition and processing unit 10, is used to establish a structured database and store the data obtained by the data acquisition and processing unit 10;

[0112] A prediction model building unit 30, connected to the database building unit 20, for building a variety of prediction models using the data stored in the database;

[0113] A model synthesis unit 40, connected to the prediction model establishment unit 30, is used to optimize each prediction model established by the prediction model establishment unit 30, and then synthesize all the optimized prediction models into a final prediction model;

[0114] The prediction unit 50 is connected to the data acquisition and processing unit 10 and the model synthesis unit 40 respectively, and obtains the sand production situation of the well to be predicted by using the data of the well to be predicted obtained by the data acquisition and processing unit 10 and the final prediction model obtained by the model synthesis unit 40.

[0115] The database established by the database establishing unit 20 is a database with a two-dimensional table structure.

[0116] The prediction models established by the prediction model establishment unit 30 include: a Bayesian classification model, a random forest model, a support vector machine model and an artificial neural network model.

[0117] From the perspective of data science, the present invention proposes a research method and prediction process for the correlation between the sand production risk of loose sandstone heavy oil layers and well location and wellbore information, reservoir physical properties, geomechanical characteristics, drilling and completion parameters, production parameters and injection and production parameters. First, the relevant data of 594 wells in the study area are collected and cleaned, and then an Access structured database for oil and gas field sand production prediction is established and improved. A full set of geological and engineering data of 250 wells in 5 oil fields is comprehensively managed to provide data for the research of sand production prediction methods based on machine learning; then, a mapping between sand production risk and geographical / reservoir physical / engineering parameters is established using a variety of machine learning algorithms such as Bayesian classification, random forest, support vector machine and artificial neural network, and a comprehensive sand production prediction and risk assessment process for loose sandstone heavy oil layers using machine learning algorithms is established. In order to improve the reliability of model training, the present invention effectively expands the database capacity and improves the prediction accuracy of the model through the "leave one out" method of cross-validation. This new comprehensive sand production risk prediction method based on machine learning technology and drilling, completion, production and injection parameter evaluation strategy provide strong guarantees for sand production risk assessment and rapid business decision-making in loose sandstone heavy oil reservoirs.

[0118] It should be noted that the above technical solution is only one implementation method of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the application methods and principles disclosed in the present invention, and it is not limited to the method described in the above specific implementation method of the present invention. Therefore, the method described above is only preferred and does not have a restrictive meaning.

Claims

1. A sand production risk prediction method, characterized by: The method comprises: Step 1: Collect and clean the data in the target area; Step 2: Establish a structured database to store the data obtained in the first step; Step 3: Use the data stored in step 2 to establish multiple prediction models; Step 4: Optimize each prediction model separately, and then combine all optimized prediction models into the final prediction model; Step 5: Use the final prediction model to obtain the sand production situation of the well to be predicted; The first step includes: Identify target areas; Collect data from each well in the target area; Perform data completion and data expansion on the collected data; The data of each well in the target area includes independent variables and dependent variables; The independent variables include the following eight types of data: The first type of data is well location geographic information, including: well location coordinates, reservoir depth, and wellbore dimensions; The second type of data is well logging data, including: acoustic wave time difference, density, resistivity and natural potential; The third type of data is reservoir characteristics, including: temperature, direction of minimum horizontal stress; The fourth type of data is reservoir geomechanical characteristics, including: overburden pressure, pore pressure, maximum and minimum horizontal in-situ stress, and minimum horizontal in-situ stress orientation; The fifth type of data is production parameters, including: daily oil production, daily water production, flow density, fluid viscosity, solution gas-oil ratio, bottom hole flow pressure and wellhead oil pressure; The sixth category of data is completion parameters, including: perforation section length, perforation section depth, perforation density, phase angle, median particle size of gravel pack completion, seam width of wire-wound screen completion, seam width of metal wool sand filter and seam width of slotted screen; The seventh type of data is steam injection parameters, including: cumulative gas injection volume, average steam injection temperature and average steam injection pressure; The eighth type of data is wellbore trajectory, including: well inclination and azimuth information of horizontal wells; The dependent variable is the sand production of a single well at a certain time point; The operations of establishing a structured database in the second step include: A database with a two-dimensional table structure is used to store all the independent variables and dependent variables obtained in the first step; The independent and dependent variables of each well are considered as a sample; All samples are divided into two parts, one is training samples and the other is validation samples; The multiple prediction models in the third step include: Bayesian classification model, random forest model, support vector machine model and artificial neural network model; The fourth step includes: Firstly, each prediction model is optimized using training samples and validation samples to obtain each optimized prediction model; Then, the model stacking method is used to iteratively process each optimized prediction model to obtain the final prediction model.

2. The sand production risk prediction method according to claim 1, characterized in that: The operation of the fifth step includes: Collect data of the well to be predicted, and process it according to the method in the first step to obtain the dependent variable of the well to be predicted; The dependent variables of the well to be predicted are input into the final prediction model, and the final prediction model outputs the sand production situation of the well to be predicted.

3. A sand production risk prediction system, characterized by: The system comprises: A data acquisition and processing unit, used to collect and clean the data of the target area; A database establishment unit, connected to the data acquisition and processing unit, is used to establish a structured database and store the data obtained by the data acquisition and processing unit; A prediction model building unit, connected to the database building unit, for building a variety of prediction models using the data stored in the database; A model synthesis unit, connected to the prediction model establishment unit, is used to optimize each prediction model established by the prediction model establishment unit, and then synthesize all the optimized prediction models into a final prediction model; The prediction unit is connected to the data acquisition and processing unit and the model synthesis unit respectively, and obtains the sand production situation of the well to be predicted by using the data of the well to be predicted obtained by the data acquisition and processing unit and the final prediction model obtained by the model synthesis unit; The data acquisition and processing unit performs the following operations: determining the target area; acquiring data of each well in the target area; completing and expanding the acquired data; the data of each well in the target area includes independent variables and dependent variables; The independent variables include the following eight types of data: The first type of data is well location geographic information, including: well location coordinates, reservoir depth, and wellbore dimensions; The second type of data is well logging data, including: acoustic wave time difference, density, resistivity and natural potential; The third type of data is reservoir characteristics, including: temperature, direction of minimum horizontal stress; The fourth type of data is reservoir geomechanical characteristics, including: overburden pressure, pore pressure, maximum and minimum horizontal in-situ stress, and minimum horizontal in-situ stress orientation; The fifth type of data is production parameters, including: daily oil production, daily water production, flow density, fluid viscosity, solution gas-oil ratio, bottom hole flow pressure and wellhead oil pressure; The sixth category of data is completion parameters, including: perforation section length, perforation section depth, perforation density, phase angle, median particle size of gravel pack completion, seam width of wire-wound screen completion, seam width of metal wool sand filter and seam width of slotted screen; The seventh type of data is steam injection parameters, including: cumulative gas injection volume, average steam injection temperature and average steam injection pressure; The eighth type of data is wellbore trajectory, including: well inclination and azimuth information of horizontal wells; The dependent variable is the sand production of a single well at a certain time point; The database establishment unit performs the following operations: A database with a two-dimensional table structure is used to store all the independent variables and dependent variables obtained in the first step; The independent and dependent variables of each well are considered as a sample; All samples are divided into two parts, one is training samples and the other is validation samples; The prediction models established by the prediction model establishment unit include: Bayesian classification model, random forest model, support vector machine model and artificial neural network model; The model synthesis unit performs the following operations: Firstly, each prediction model is optimized using training samples and validation samples to obtain each optimized prediction model; Then, the model stacking method is used to iteratively process each optimized prediction model to obtain the final prediction model.

Images