A prediction method for daily oil production and water content of oil wells based on deep learning

Abstract

The present invention discloses a method for predicting daily oil production and water content of oil wells based on deep learning, including S1, establishing a learning sample selection standard, and selecting data that meets the standard from the historical data of the target oil field as a learning sample; S2, analyzing the many factors affecting the development of the oil field based on the learning sample, and then screening out the index data that meets the requirements, and forming a single well production and water content impact index system; S3, performing data preprocessing on the index data that meets the requirements to form a data set for model training; S4, building, training and verifying a hybrid deep learning network model for predicting daily oil production and water content of oil wells based on the data set; S5, using the trained hybrid deep learning network model to predict the daily oil production and water content of the oil well. The advantage is that it can significantly improve the accuracy and stability of the prediction of daily oil production and water content of the oil well.

Description

Technical Field

[0001] The present invention relates to the technical field of oilfield development, and in particular to a method for predicting daily oil production and water content of an oil well based on deep learning. Background Art

[0002] At present, water injection development is the main method of oil field development in China, and its output accounts for 80% of crude oil production. Among them, the prediction of single well production and water content change law is the basis for oil field production profile prediction, and plays an important role in investment decision-making and other work. This work runs through the entire development process of water-driven oil fields. The prediction methods used in the early days include oil testing, analogy method, empirical method, water-driven oil experiment and other methods, but due to the small number of wells and poor data representativeness, the prediction results vary greatly. At present, the commonly used prediction methods include numerical simulation, decline curve, water-driven curve and other methods. The prediction methods are relatively reliable, but they all have certain limitations. They require oil field production to remain relatively stable without major adjustments. In addition, due to the long production history of oil wells and the large amount of data, more manpower and material resources are required for calculation, which is time-consuming and requires high professional skills of personnel.

[0003] In recent years, with the breakthrough and development of artificial intelligence technology, deep learning in artificial intelligence has gradually been applied to various fields. Deep learning is a technology that can learn experience from data, and establish complex models through powerful nonlinear mapping capabilities to summarize production laws and predict the future. In the field of oilfield development, a large number of surveys have also shown that research work has been carried out in exploration, drilling, production, and reservoirs. However, in terms of oil well production and water content prediction, it is in its infancy. There is a lack of artificial intelligence methods to predict production and water content. The influencing index system, sample selection criteria, algorithm optimization principles, and model optimization methods. Therefore, it is of great significance to develop a prediction method for multi-well daily oil production and water content based on deep learning for intelligent oilfield production allocation. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting daily oil production and water content of oil wells based on deep learning, so as to solve the above-mentioned problems existing in the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting daily oil production and water content of an oil well based on deep learning, comprising the following steps:

[0007] S1. Establish a learning sample selection standard and select data that meets the standard from the historical data of the target oil field as learning samples;

[0008] S2. Analyze the many factors that affect oilfield development based on the learning samples, and then select the index data that meet the requirements, and form a single well production and water content impact index system;

[0009] S3. Preprocess the indicator data that meets the requirements to form a data set for model training;

[0010] S4. Build, train and validate a hybrid deep learning network model for predicting daily oil production and water content of oil wells based on the dataset;

[0011] S5. Use the trained hybrid deep learning network model to predict the daily oil production and water content of the oil well.

[0012] Preferably, step S1 specifically includes performing data volume analysis and evaluation on the historical data of the target oil field, formulating learning sample selection criteria based on whether the historical data of each pilot block in the target oil field meets the needs of research and modeling, and screening out oil wells and historical data that meet the learning sample selection criteria from the historical data of the target oil field as learning samples.

[0013] Preferably, step S2 specifically includes the following contents:

[0014] S21. Through qualitative analysis, we preliminarily analyze the many factors that affect oilfield development, and then fully consider the development characteristics and on-site measures, and then determine the five major indicators that affect oil well production and water content, namely, well basic data, reservoir basic data, logging interpretation results data, oil well production and status data, and oil well measures effect data;

[0015] S22. According to the physical meaning of production and water content and influencing indicators, analyze the logical relationship between indicators and eliminate some influencing indicators; calculate the variance of each indicator item. Indicators with a variance of 0 have the same impact on all oil wells in this pilot block, and some indicators are eliminated again; and the remaining indicators are retained;

[0016] S23. Use Pearson correlation analysis to quantitatively analyze the retained indicators and form a single well production and water content impact indicator system.

[0017] Preferably, the qualitative analysis in step S21 is specifically to sort out the many factors that affect oil field development based on seepage theory, water injection development characteristics of medium- and high-permeability sandstone reservoirs, and production history.

[0018] Preferably, step S3 specifically includes performing quality statistical analysis on the indicator data after quantitative analysis, and assigning corresponding data preprocessing methods to perform data preprocessing according to the actual situation and business rules of the collected indicator data.

[0019] Preferably, the data preprocessing method includes outlier processing, missing value processing, data type conversion, and data normalization.

[0020] Preferably, step S4 specifically includes the following contents:

[0021] S41, dividing the preprocessed index data into a test set and a training set according to a certain ratio;

[0022] S42. Using an artificial neural network algorithm tool to build a hybrid deep learning network model, wherein the hybrid deep learning network model is based on a complementary combination of a convolutional neural network and a bidirectional long short-term memory network;

[0023] S43, using the training set and the test set to train and test the hybrid deep learning network model respectively; and using MSE and MAE as error evaluation criteria, when the error evaluation criteria between the prediction results of the training set and the prediction results of the test set are the smallest, obtaining the trained hybrid deep learning network model;

[0024] S44. Perform ten-fold cross validation on the trained hybrid deep learning network model to obtain the corresponding accuracy of each validation; the average accuracy of the 10 results is used as an estimate of the accuracy of the algorithm, and then perform multiple ten-fold cross validations to calculate the average value as the accuracy of the trained hybrid deep learning network model.

[0025] Preferably, in step S41, since the feature dimension is too large, the training set is subjected to dimensionality reduction processing, and the principal component analysis method is used to transform the multiple indicators into three comprehensive indicators to extract the main feature components of the data.

[0026] Preferably, the construction process of the hybrid deep learning network model in step S42 is to first input the input data into the bidirectional long short-term memory network, use the softsign activation function, train the sample data through the forward and backward networks, and then send the output data into the one-dimensional convolutional neural network to effectively fit the input and output data, and finally build the hybrid deep learning network model; the softsign activation function formula is,

[0027]

[0028] Preferably, step S5 specifically includes: inputting the historical data of the oil well into the trained hybrid deep learning network model after data preprocessing, predicting the daily oil production and water content of the oil well in the next month, and repeating the above prediction process until the prediction within a certain production time is completed.

[0029] The beneficial effects of the present invention are as follows: 1. In terms of data processing, the decline of oil wells, the law of water content increase, and the long-term production history data of oil wells in different development stages are analyzed and processed, and a credible and scientific sample selection standard is established to provide a basis for machine learning. Qualitative analysis, logical analysis, and quantitative analysis are adopted, and the dynamic and static factors of oil field development are comprehensively considered to construct an index system affecting daily oil production and water content, so that the learning samples in terms of oil well production and water content prediction are more representative. 2. In terms of model construction, Bi-LSTM and convolutional neural network CNN are combined, and a hybrid deep learning model is used for prediction. The commonly used model construction method in the oil field field is a double-layer LSTM. The hybrid model has the characteristics of fast training time, small number of cycles required when the model reaches stability, and high model accuracy, and can predict daily oil production and water content faster and more accurately. In addition, for LSTM, the activation function uses softsign instead of traditional softmax, tanh and other activation functions, making the model faster and less prone to saturation (about 0 gradient). 3. Compared with the existing prediction methods of daily oil production and water content, the method of the present invention can significantly improve the accuracy and stability of the prediction of daily oil production and water content of oil wells. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a logic flow chart of a prediction method in an embodiment of the present invention;

[0031] Figure 2 This is a quantitative analysis result diagram using Pearson correlation analysis in an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of a hybrid deep learning network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] like Figure 1 As shown, in this embodiment, a method for predicting daily oil production and water content of an oil well based on deep learning is provided, comprising the following steps:

[0035] S1. Establish a learning sample selection standard and select data that meets the standard from the historical data of the target oil field as learning samples;

[0036] S2. Analyze the many factors that affect oilfield development based on the learning samples, and then select the index data that meet the requirements, and form a single well production and water content impact index system;

[0037] S3. Preprocess the indicator data that meets the requirements to form a data set for model training;

[0038] S4. Build, train and validate a hybrid deep learning network model for predicting daily oil production and water content of oil wells based on the dataset;

[0039] S5. Use the trained hybrid deep learning network model to predict the daily oil production and water content of the oil well.

[0040] In summary, the method of the present invention mainly includes five parts, namely, obtaining learning samples, constructing a single well production and water content impact index system, data preprocessing, model construction and training verification, and daily oil production and water content prediction. The following will explain these five parts separately.

[0041] 1. Obtaining learning samples

[0042] This part corresponds to step S1, specifically, analyzing and evaluating the data volume of the historical data of the target oil field, formulating learning sample selection criteria based on whether the historical data of each pilot block in the target oil field meets the needs of research and modeling, and screening out oil wells and historical data that meet the learning sample selection criteria from the historical data of the target oil field as learning samples.

[0043] For example, the selection criteria for production wells: mechanical oil production, vertical wells; the selection criteria for production stages: high water content (fw>60%), natural decline stage, continuous shut-in less than 6 months. Finally, 551 wells and 25,384 learning samples were determined.

[0044] 2. Constructing a single well production and water content impact index system

[0045] This part corresponds to step S2, and specifically includes the following contents:

[0046] S21. Through qualitative analysis, we preliminarily analyzed the many factors that affect oil field development (geological factors, development factors, and measures factors), and then fully considered the development characteristics and on-site measures (including development system adjustment, well network densification, measures workload, etc.), and then determined the five major categories of indicators that affect oil well production and water content, including well basic data, reservoir basic data, logging interpretation results data, oil well production and status data, and oil well measures effect data, totaling 64 indicators.

[0047] Among them, the qualitative analysis is to sort out the many factors that affect oil field development based on seepage theory, water injection development characteristics of medium and high permeability sandstone reservoirs, and production history.

[0048] S22. According to the physical meaning of production and water content and influencing indicators, analyze the logical relationship between indicators (including cause-effect relationship, equivalent relationship, process relationship), and eliminate some (28 items) influencing indicators; calculate the variance of each indicator item. Indicators with a variance of 0 have the same impact on all oil wells in this pilot block, and eliminate some indicators again, specifically 10 indicators such as initial formation temperature. After the above logical analysis, 25 indicators are retained, including daily oil production, water content, horizontal axis, vertical axis, effective thickness, permeability, well spacing, static pressure, flow pressure, dynamic liquid level, stroke, stroke times, monthly oil production, monthly liquid production, cumulative oil production, pump diameter, pump depth, oil nozzle, cumulative liquid production, formation pressure, degree of recovery, daily liquid production, daily oil production last month, water content last month, and daily liquid production last month;

[0049] S23, use Pearson correlation analysis to quantitatively analyze the retained indicators, such as Figure 2 As shown, the analysis results are consistent with the analysis results of step S22. Oil production is proportional to well spacing, effective thickness, flowing pressure, etc., and inversely proportional to recovery degree, water content, etc. The results of quantitative analysis provide a basis for subsequent feature selection, feature extraction and feature construction, and ultimately form a single well production and water content impact index system.

[0050] 3. Data Preprocessing

[0051] This part corresponds to step S3, specifically, performing quality statistical analysis on the indicator data after quantitative analysis, and assigning corresponding data preprocessing methods to perform data preprocessing according to the actual situation and business rules of the collected indicator data.

[0052] In this embodiment, the data preprocessing method includes the steps of outlier processing, missing value processing, data type conversion, data normalization, etc. Specifically:

[0053] The missing parameter data is filled by interpolation, and the smoothing method is used to solve the problem of large oil production gap between two months in a year. The data is processed by median filtering to solve the problem of serious uneven oil production in a well every year. Finally, a consistency test is performed to check the coordination between the importance of each element.

[0054] 4. Model Construction and Training Verification

[0055] This part corresponds to step S4, and specifically includes the following contents:

[0056] S41, the preprocessed index data is divided into a test set and a training set according to a certain ratio; because the feature dimension is too large, the training set is reduced in dimension, and the principal component analysis method is used to transform multiple indicators into three comprehensive indicators to extract the main characteristic components of the data. At this point, all data preparation work is completed.

[0057] S42. According to the processed data, a hybrid deep learning network model is designed. The hybrid deep learning network model is based on the complementary combination of convolutional neural network (CNN) and bidirectional long short-term memory network (Bi-LSTM), such as Figure 3 As shown. The model building process is: using open source mature artificial neural network algorithm tool modules to build a hybrid deep learning network model. First, input the input data into the bidirectional long short-term memory network (Bi-LSTM), use the softsign activation function, train the sample data through the forward and backward networks, and then send the output data to the one-dimensional convolutional neural network to effectively fit the input and output data, and finally build a hybrid deep learning network model. The softsign formula is:

[0058]

[0059] S43, using the training set and the test set to train and test the hybrid deep learning network model respectively; and using MSE and MAE as error evaluation criteria, when the error evaluation criteria between the prediction results of the training set and the prediction results of the test set are the smallest, obtaining the trained hybrid deep learning network model;

[0060] S44. Perform ten-fold cross validation on the trained hybrid deep learning network model to obtain the corresponding accuracy of each validation; the average accuracy of the 10 results is used as an estimate of the accuracy of the algorithm, and then perform multiple ten-fold cross validations to calculate the average value as the accuracy of the trained hybrid deep learning network model.

[0061] The specific method of 10-fold cross validation of the model is to divide the sample data set into 10 parts, take 9 of them as training data and 1 as test data in turn, and conduct experiments to obtain the corresponding accuracy. The average accuracy of the 10 results is used as an estimate of the accuracy of the algorithm, and then multiple 10-fold cross validations are performed to calculate the average, and finally the model accuracy is determined to be 95%.

[0062] 5. Daily oil production and water content prediction

[0063] This part corresponds to step S5, specifically, preparing prediction data according to the format requirements of the feature data provided in step 41, and inputting the trained hybrid deep learning network model after data preprocessing to predict the daily oil production and water content of the oil well in the next month. Repeat the above prediction process until the prediction within a certain production time is completed.

[0064] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:

[0065] The present invention provides a method for predicting the water content of daily oil well production based on deep learning. In terms of data processing, the method analyzes and processes the oil well decline, water content increase law, and long-term production history data of oil wells in different development stages, establishes a credible and scientific sample selection standard, and provides a basis for machine learning. Qualitative analysis, logical analysis, and quantitative analysis are adopted to comprehensively consider the dynamic and static factors of oil field development, and construct an index system affecting daily oil production and water content, so that the learning samples in terms of oil well production and water content prediction are more representative. In terms of model construction, the method combines Bi-LSTM and convolutional neural network CNN, and uses a hybrid deep learning model for prediction. The commonly used model construction method in the oil field field is a double-layer LSTM. The hybrid model has the characteristics of fast training time, small number of cycles required when the model reaches stability, and high model accuracy, and can predict daily oil production and water content faster and more accurately. In addition, for LSTM, the activation function uses softsign instead of traditional softmax, tanh and other activation functions, making the model faster and less prone to saturation (about 0 gradient). Compared with the existing prediction methods of daily oil production and water content, the method of the present invention can significantly improve the accuracy and stability of the prediction of daily oil production and water content of oil wells.

[0066] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.

Claims

1. A method for predicting daily oil production and water content of oil wells based on deep learning, characterized in that: The following steps are included: S1. Establish a learning sample selection standard and select data that meets the standard from the historical data of the target oil field as learning samples; Step S1 specifically includes: analyzing and evaluating the data volume of the historical data of the target oil field, formulating a learning sample selection standard based on whether the historical data of each pilot block in the target oil field meets the needs of research and modeling, and selecting oil wells and historical data that meet the learning sample selection standard from the historical data of the target oil field as learning samples; S2. Analyze the many factors that affect oilfield development based on the learning samples, and then select the index data that meet the requirements, and form a single well production and water content impact index system; Step S2 specifically includes the following contents: S21. Through qualitative analysis, we preliminarily analyze the many factors that affect oilfield development, and then fully consider the development characteristics and on-site measures, and then determine the five major indicators that affect oil well production and water content, namely, well basic data, reservoir basic data, logging interpretation results data, oil well production and status data, and oil well measures effect data; The qualitative analysis in step S21 specifically includes sorting out the many factors that affect oilfield development based on seepage theory, water injection development characteristics of medium- and high-permeability sandstone reservoirs, and production history; S22. According to the physical meaning of production and water content and influencing indicators, analyze the logical relationship between indicators and eliminate some influencing indicators; calculate the variance of each indicator item. Indicators with a variance of 0 have the same impact on all oil wells in this pilot block, and some indicators are eliminated again; and the remaining indicators are retained; S23. Quantitatively analyze the retained indicators using Pearson correlation analysis to form a single well production and water content impact indicator system; S3. Preprocess the indicator data that meets the requirements to form a data set for model training; S4. Build, train and validate a hybrid deep learning network model for predicting daily oil production and water content of oil wells based on the dataset; S5. Use the trained hybrid deep learning network model to predict the daily oil production and water content of the oil well.

2. The method for predicting daily oil production and water content of an oil well based on deep learning according to claim 1, characterized in that: Step S3 specifically includes performing quality statistical analysis on the indicator data after quantitative analysis, and assigning corresponding data preprocessing methods to perform data preprocessing according to the actual situation and business rules of the collected indicator data.

3. The method for predicting daily oil production and water content of an oil well based on deep learning according to claim 2, characterized in that: The data preprocessing method includes outlier processing, missing value processing, data type conversion, and data normalization.

4. The method for predicting daily oil production and water content of an oil well based on deep learning according to claim 1, characterized in that: Step S4 specifically includes the following contents: S41, dividing the preprocessed index data into a test set and a training set according to a certain ratio; S42. Using an artificial neural network algorithm tool to build a hybrid deep learning network model, wherein the hybrid deep learning network model is based on a complementary combination of a convolutional neural network and a bidirectional long short-term memory network; S43, using the training set and the test set to train and test the hybrid deep learning network model respectively; and using MSE and MAE as error evaluation criteria, when the error evaluation criteria between the prediction results of the training set and the prediction results of the test set are the smallest, obtaining the trained hybrid deep learning network model; S44. Perform ten-fold cross validation on the trained hybrid deep learning network model to obtain the corresponding accuracy of each validation; the average accuracy of the 10 results is used as an estimate of the accuracy of the algorithm, and then perform multiple ten-fold cross validations to calculate the average value as the accuracy of the trained hybrid deep learning network model.

5. The method for predicting daily oil production and water content of an oil well based on deep learning according to claim 4, characterized in that: In step S41, since the feature dimension is too large, the training set is subjected to dimensionality reduction processing, and the principal component analysis method is used to transform multiple indicators into three comprehensive indicators to extract the main feature components of the data.

6. The method for predicting daily oil production and water content of an oil well based on deep learning according to claim 4, characterized in that: The construction process of the hybrid deep learning network model in step S42 is to first input the input data into the bidirectional long short-term memory network, use the softsign activation function, train the sample data through the forward and backward networks, and then send the output data into the one-dimensional convolutional neural network to effectively fit the input and output data, and finally build the hybrid deep learning network model; the softsign activation function formula is, 7. The method for predicting daily oil production and water content of an oil well based on deep learning according to claim 1, characterized in that: Step S5 specifically includes: inserting the historical data of the oil well into the trained hybrid deep learning network model after data preprocessing, predicting the daily oil production and water content of the oil well in the next month, and repeating the above prediction process until the prediction within a certain production time is completed.

Images