Water-driven oil field single well yield prediction method fusing spatio-temporal information

Abstract

The invention provides a water-drive oil field single well yield prediction method fusing spatio-temporal information. The water-drive oil field single well yield prediction method fusing spatio-temporal information comprises the steps that 1, the importance of influence factors is evaluated through an embedded method; 2, screening and determining dynamic and static influence factors by adopting a voting method; 3, determining spatial dynamic influence factors based on sequence correlation analysis; 4, constructing a water injection composite sample considering the response hysteresis; 5, constructing a single well development index prediction model based on LSTM and MLP, wherein the single well development index prediction model integrates time-space dynamic and static information; and 6, carrying out model prediction. According to the water drive oil field single well yield prediction method fusing the spatio-temporal information, development index prediction fusing mechanism guidance and data driving is achieved, the prediction precision is high, the generalization ability is high, and a new means is provided for prediction of the oil field single well yield.

Description

Technical Field

[0001] The present invention relates to the technical field of oilfield time-series data processing, and particularly to a method for predicting the single-well production of a waterflooded oilfield by integrating spatio-temporal information. Background Art

[0002] Studying the variation law of the single-well production of an oilfield and accurately predicting the single-well production are helpful for guiding the formulation of oilfield development plans and the adjustment of development directions, and are of great significance for the efficient development of oilfields. Predicting the single-well production has always been a difficult problem. Traditional prediction methods are either various trend line fitting methods based on time series or reservoir numerical simulation methods based on the numerical solutions of partial differential equations. The former method is simple, but it is no longer applicable in the middle and late stages of development when the oil-water changes are complex, and it is impossible to effectively analyze the influence of factors on indicators from the data. The latter is more accurate and can also effectively analyze the influence of factors on indicators, but it requires a lot of data, a large amount of work, is time-consuming and laborious, and is not conducive to real-time guiding the scheme adjustment.

[0003] At present, the digital and information levels of major domestic oilfields are constantly improving, and a large amount of exploration and production data have been accumulated. Therefore, using data mining technology to predict development indicators has become a research hotspot. Traditional ideas include using methods such as neural networks, support vector machines (SVMs), logistic regression, and random forests to train and predict samples. However, such methods have a basic requirement that the samples should be independent of each other. However, the dynamic factors and development indicators of a single well are all time-series data and do not meet the independence requirement, resulting in a general inability to meet the on-site requirements in terms of generalization ability although the training effect is good when using such methods. Another idea is to use the long short-term memory network model (LSTM), which can extract the change characteristics of dynamic time-series factors and combine them with static information (such as permeability, effective thickness, etc.) to achieve a high-precision fitting of historical development dynamics and has achieved good results.

[0004] However, the development effect of a single well is not only related to static information and historical development dynamics, but also affected by the water injection of adjacent water wells. That is to say, to accurately predict the development indicators of old wells in the middle and late stages, in addition to considering the influence of its static and dynamic factors, it is also necessary to consider the water injection effectiveness in space. However, at present, there is a lack of an effective method for comprehensively considering the influence of oil well static parameters (static), development dynamic factors (time), and adjacent water wells (space) on single-well production prediction.

[0005] In the Chinese patent application with the application number: CN201910122231.2, it involves a method and device for predicting the liquid production capacity of horizontal wells in stress-sensitive reservoirs, including: obtaining the static parameters of the reservoir and fluid and the relevant production parameters of the studied oil well; determining the initial equivalent permeability of the fractured reservoir according to the static parameters; based on the oil-gas two-phase inflow dynamic equation and the pure water-phase inflow dynamic equation of the horizontal well in the stress-sensitive reservoir, determining the oil production and water production at different bottom-hole flowing pressures respectively according to the static parameters, relevant production parameters and initial equivalent permeability; determining the liquid production at different bottom-hole flowing pressures according to the oil production and water production at different bottom-hole flowing pressures; predicting the liquid production capacity of the studied oil well based on the liquid production at different bottom-hole flowing pressures. Considering the stress sensitivity of the reservoir, the change of the average reservoir pressure, and the actual situation of the co-production of oil, gas and water near the bottom hole, it can more accurately predict the liquid production capacity of horizontal wells in fractured reservoirs and provide support for the efficient development of such reservoirs.

[0006] In the Chinese patent application with the application number: CN201710706050.5, it involves a method for predicting the development indexes of steam-assisted gravity drainage in a new development area, including: classifying the adjacent developed areas according to the static characteristic parameters; selecting several typical wells in the static characteristic categories and establishing numerical mechanism models respectively; selecting the development characteristic parameters and using the numerical mechanism models to solve the development characteristic parameters of each typical well; calculating the correction coefficient between the development characteristic parameters solved by the numerical mechanism models and the actual production; classifying the new development area with the same static characteristic parameter categories and establishing the numerical mechanism models of single-well pairs according to the relevant static parameters in the range where the planned well positions in the new development area are located; calculating the development characteristic parameters of each single-well pair under different static characteristic parameter categories; correcting the results of each development characteristic parameter; obtaining the corrected development indexes of single wells in the new development area; obtaining the overall development indexes of the new development area.

[0007] In the Chinese patent application with the application number: CN202010641936.8, it involves a method for predicting the recoverable reserves and evaluating the development effect of a top-gas and edge-water reservoir. This method first starts from the reservoir characteristics and development mode to establish an oil reservoir engineering method model for the top-gas and edge-water reservoir; secondly, starting from the seepage theory, it deduces and establishes a method for evaluating the gas and water drive production contributions of the top-gas and edge-water reservoir, and evaluates the cumulative oil production of the gas drive part and the water drive part of such reservoirs; finally, based on the evaluation results of the production contributions, using the gas drive characteristic curve and the water drive characteristic curve, a method for predicting the recoverable reserves and evaluating the development effect of the top-gas and edge-water reservoir is established in two aspects: the gas drive part and the water drive part. The method of the present invention provides a convenient method for predicting the recoverable reserves and evaluating the development effect in the middle and late stages of the development of the top-gas and edge-water reservoir, and provides a basis for the optimization of oilfield measures and the identification of remaining oil.

[0008] The above existing technologies are quite different from the present invention and cannot solve the technical problems we want to address. Therefore, we have invented a new method for predicting the single-well production of waterflooding oilfields by integrating spatio-temporal information. Summary of the Invention

[0009] The objective of the present invention is to provide a method for predicting development indicators that integrates mechanism guidance and data-driven approaches, with high prediction accuracy and strong generalization ability, for predicting the single-well production of waterflooding oilfields by integrating spatio-temporal information.

[0010] The objective of the present invention can be achieved through the following technical measures: A method for predicting the single-well production of waterflooding oilfields by integrating spatio-temporal information, which includes:

[0011] Step 1: Evaluate the importance of influencing factors using an embedded method;

[0012] Step 2: Use the voting method to screen and determine dynamic and static influencing factors;

[0013] Step 3: Determine the spatial dynamic influencing factors based on sequence correlation analysis;

[0014] Step 4: Construct an injection composite sample considering the lag effect of water injection;

[0015] Step 5: Construct a single-well development index prediction model that integrates spatio-temporal dynamic and static information based on LSTM and MLP;

[0016] Step 6: Conduct model prediction.

[0017] The objective of the present invention can also be achieved through the following technical measures:

[0018] In Step 1, for the dynamic and static influencing factors that affect the single-well production effect, evaluate the importance of the influencing factors using an embedded method.

[0019] In Step 1, use three algorithms based on tree models, namely decision tree, random forest, and Xgboost methods, to obtain the importance ranking of each influencing factor respectively.

[0020] In Step 2, use the weighted voting method for screening. The principle of the weighted voting method is to assign a weight reflecting its overall performance to each individual learner, and measure its overall performance based on the validation set, so that the votes of more accurate models will be effectively counted multiple times.

[0021] In Step 2, the formula of the weighted voting method is expressed as follows:

[0022]

[0023] Among them,

[0024] where: x is the input test sample; ω i is the weight of classifier c i in the present invention, it is one third; represents the probability that the base classifier i predicts that x belongs to class j; j * is the maximum probability that the voting classifier finally predicts that x belongs to class j.

[0025] In step 3, the significantly relevant water injection time series, including spatial and dynamic factors, are taken as the factors affecting the well dynamics; due to the existence of the spatial well spacing between oil wells and water wells, the dynamics of the water wells are reflected in the oil wells with a certain time lag.

[0026] In step 3, for the oil wells whose production is to be predicted, according to the geodetic coordinates of the oil wells, the well numbers of the water wells within a radius of 500 meters of the oil wells are screened in the database through an algorithm to obtain the corresponding relationship table between the oil wells and their adjacent water wells; at this time, one oil well may correspond to multiple water wells.

[0027] In step 3, in the oil well and the corresponding water well table, the time series of the liquid production of this oil well and the water injection volumes of all corresponding water wells are extracted according to the well numbers, and the serial correlation analysis of different lag times is carried out for each water well to determine the correlation coefficients of different lag time series of each water well.

[0028] In step 3, select the water injection wells with a lag order < 12 and the absolute value of the correlation coefficient being the largest and greater than 0.6 as the corresponding water wells of the oil well; the lag order at which the absolute value of the correlation coefficient is the largest is m; for the oil wells where the threshold conditions are not met for the nearby water injection wells, that is, the situation where the water injection effectiveness cannot be clearly determined, the average value series of the injection volumes of these water wells are inserted into the input factors to unify the input sequence format.

[0029] In step 4, for all oil wells, taking the latest time node of the training data as the reference point, start intercepting the oil production data of the target oil well for 60 months backwards as the output, and the corresponding other dynamic factor data as the input; for each adjacent water injection well i corresponding to an oil well, according to the lag order m when the correlation coefficient is the largest that has been determined, start intercepting the water injection well injection volume data for 60 months m months backwards on the basis of the time period corresponding to the oil well; analyze all the oil wells and water wells in a block or an oilfield to construct a composite sample.

[0030] In step 5, based on the multi-layer perceptron network MLP and the LSTM network, construct a machine learning network model for the fusion of dynamic and static information; the output formula of the machine learning model obtained by weighted synthesis of the neuron values after being processed by the MLP network and the LSTM network is as follows:

[0031]

[0032] Among them, X i is the value of the i-th output neuron processed by the MLP neural network for static features; Y i represents the value of the i-th neuron output by the LSTM neural network for dynamic factors; Z i represents the value of the i-th neuron output by the LSTM neural network for the water injection volume of adjacent wells; K i represents the i-th convolution kernel, * represents convolution, c represents the number of channels, Ψ() is the response function of the fully connected layer, taken as the rectified linear unit ReLu; the overall training of the model uses the adaptive moment estimation Adam algorithm.

[0033] In step 5, the samples for model construction are not limited to a single well, but all single well samples after spatio-temporal fusion are substituted into the above model for modeling; the data samples are divided into training and testing according to a certain ratio, the training algorithm uses the adaptive moment estimation Adam algorithm, the loss functions use the mean square error mse and the coefficient of determination r2, and by selecting an appropriate batch size and number of training times, the optimal learning model can be obtained.

[0034] In step 6, the influence of the change in the monthly working system on the development effect is considered among the influencing factors; during prediction, if the change in the working system can be given artificially, this change can be directly input into the model as an influencing factor; if there is no change in the working system, prediction for multiple future steps is carried out by the method of rolling completion.

[0035] In step 6, during completion, since there is no change in static factors, they are directly completed; dynamic factors include the stroke and the number of strokes per minute, and the moving average method is used for completion and used as the input of the model.

[0036] The object of the present invention can also be achieved by the following technical measures: a waterflooding oilfield single well production prediction system that fuses spatio-temporal information, and this waterflooding oilfield single well production prediction system that fuses spatio-temporal information uses the waterflooding oilfield single well production prediction method that fuses spatio-temporal information to predict the waterflooding oilfield single well production.

[0037] The single-well production prediction method for waterflooding oilfields integrating spatio-temporal information in the present invention comprehensively considers the influence of the static characteristics, dynamic characteristics of a single well and the change trend of the characteristics of spatially adjacent water wells on the single-well production prediction. By analyzing the correlation relationship between the oil well and the adjacent water wells through the time series method, it realizes obtaining the injection-production correspondence relationship in oil production engineering by means of data mining, and constructs training samples integrating static, dynamic and spatial information based on this. Using a machine learning model for training and prediction. In this process, the framework of the machine learning model simulates the basic process of development dynamic analysis, realizes the prediction of development indicators integrating mechanism guidance and data-driven, has high prediction accuracy and strong generalization ability, and provides a new means for predicting the single-well production of oilfields. The present invention analyzes the sequence correlation between adjacent injection wells and production wells, constructs influence samples of water injection volume according to the correlation coefficient and the lag order; inputs static conditions into the MLP neural network module, dynamic conditions into the LSTM neural network module, and water injection samples into the LSTM neural network module, and uses a fully connected layer to connect the three modules in parallel and output the prediction result; this method comprehensively considers the connection between the static characteristics, the change trend of dynamic data and the influence of injection wells in space on the single-well development indicators to be predicted, and realizes the integration of mechanism guidance and data-driven. The effect of this model has been greatly improved compared with general machine learning models, and it can effectively predict the single-well production of oilfields. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flowchart of a specific embodiment of the single-well production prediction method for waterflooding oilfields integrating spatio-temporal information of the present invention;

[0039] Figure 2 It is a schematic diagram of the importance ranking of three embedded features in a specific embodiment of the present invention;

[0040] Figure 3 It is a sequence correlation diagram of injection-production data between adjacent injection wells and production wells in a specific embodiment of the present invention;

[0041] Figure 4 It is a diagram of the composite sample construction method considering water injection in a specific embodiment of the present invention;

[0042] Figure 5 It is a diagram of the model training process with a time window of 12 in a specific embodiment of the present invention;

[0043] Figure 6 It is a diagram of the model fitting result with a time window of 12 in a specific embodiment of the present invention;

[0044] Figure 7 It is a diagram of the model prediction result with a time window of 12 in a specific embodiment of the present invention;

[0045] Figure 8Rolling prediction fitting comparison diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0046] Figure 9 Rolling prediction comparison diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0047] Figure 10 Rolling prediction result diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0048] Figure 11 Schematic diagram of the machine learning network model for fusing dynamic and static information constructed in a specific embodiment of the present invention.

[0049] Figure 12 Correlation diagram of injection-production data sequences of adjacent injection wells and production wells in a specific embodiment of the present invention;

[0050] Figure 13 Rolling prediction fitting comparison diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0051] Figure 14 Rolling prediction comparison diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0052] Figure 15 Rolling prediction result diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0053] Figure 16 Correlation diagram of injection-production data sequences of adjacent injection wells and production wells in a specific embodiment of the present invention;

[0054] Figure 17 Rolling prediction fitting comparison diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0055] Figure 18 Rolling prediction comparison diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention;

[0056] Figure 19 Rolling prediction result diagram of the prediction model based on fused spatio-temporal information in a specific embodiment of the present invention. Detailed implementation mode

[0057] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0058] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should also be understood that when the terms "comprise" and / or "include" are used in this specification, they specify the presence of features, steps, operations, and / or combinations thereof.

[0059] The present invention aims to provide a complete process for predicting the single-well production of a waterflooding oilfield. This process mainly realizes the feature extraction of static and dynamic factors of a single well and the injection-production correspondence relationship of spatially adjacent water wells, and based on this, a machine learning model for single-well production prediction is established. The calculation methods include:

[0060] Step 1: Evaluate the importance of influencing factors using an embedded method

[0061] For dynamic and static influencing factors that affect the single-well production effect, the importance of the influencing factors is evaluated using an embedded method. That is, three algorithms based on tree models, namely decision tree, random forest, and Xgboost methods, are used to obtain the importance ranking of each influencing factor respectively.

[0062] Step 2: Screen and determine dynamic and static influencing factors using the voting method

[0063] The weighted voting method is used for screening. The principle of the weighted voting method is to assign a weight reflecting its overall performance to each individual learner. It is usually measured according to the validation dataset, so that the votes of more accurate models will be effectively calculated multiple times. The formula of the weighted voting method is expressed as follows:

[0064]

[0065] Where,

[0066] In the formula: x is the input test sample; ω i is the weight of the classifier c i in the present invention is one-third. represents the probability that the base classifier i predicts that x belongs to class j. j * is the maximum probability that the voting classifier finally predicts that x belongs to class j.

[0067] Step 3: Determine the spatial dynamic influencing factors based on sequence correlation analysis

[0068] For water - drive oilfields, factors affecting the performance of oil wells not only include their own operating systems and reservoir conditions, etc. More importantly, the injection - production correspondence relationship in space and the energy replenishment situation are also very important aspects. The time series of injection volume (spatial and dynamic factors) that is significantly relevant is taken as a factor affecting the dynamics of oil wells. Due to the existence of the spatial well spacing between oil and water wells, the dynamics of water wells are reflected in oil wells with a certain time lag, which cannot be considered by conventional methods.

[0069] For the oil wells whose production needs to be predicted, according to the geodetic coordinates of the oil wells, the well numbers of water wells within a radius of 500 meters of the oil wells are screened through an algorithm in the database to obtain the corresponding relationship table between the oil wells and their adjacent water wells. At this time, one oil well may correspond to multiple water wells.

[0070] What determines the injection - production relationship between water wells and oil wells is the underground connectivity. Although the spatial distance is close, whether they are connected underground still requires data analysis of the actual development dynamics. In the table of oil wells and corresponding water wells, the time series of the liquid production volume of this oil well and the injection volumes of all corresponding water wells are extracted according to the well numbers. For each water well, sequence correlation analysis with different lag times is carried out to determine the correlation coefficients of different lag - time sequences of each water well. Select the water injection wells with a lag order < 12 and the absolute value of the correlation coefficient being the largest and greater than 0.6 as the corresponding water wells of the oil well. The lag order at which the absolute value of the correlation coefficient is the largest is m. For oil wells that do not meet the threshold conditions for all nearby water injection wells (the situation where the water injection effect cannot be clearly determined), the average value sequence of the injection volumes of these water wells is inserted into the input factors, and the unified input sequence format is input.

[0071] Step 4: Construct a water - injection composite sample considering the lag of effectiveness

[0072] For all oil wells, taking the latest time node of the training data as the reference point, start intercepting the oil production data of the target oil well for the previous 60 months as the output, and the data of other dynamic factors as the input. For each adjacent injection well i corresponding to an oil well, according to the lag order m when the correlation coefficient is the largest that has been determined, start intercepting the injection volume data of the injection well for 60 months starting m months earlier based on the time period corresponding to the oil well. Analyze all the oil and water wells in a block or an oilfield to construct a composite sample.

[0073] Step 5: Construct a single - well development index prediction model that integrates spatio - temporal dynamic and static information based on LSTM and MLP

[0074] Based on the multi - layer perceptron network (MLP) and the LSTM network, construct a machine - learning network model that integrates dynamic and static information. Its model structure is as Figure 11 shown.

[0075] The output formula of the machine learning model is obtained by weighted synthesis of the neuron values processed by the MLP network and the LSTM network as follows:

[0076]

[0077] Among them, X i is the value of the i-th output neuron processed by the MLP neural network for static features; Y i represents the value of the i-th neuron output by the LSTM neural network for dynamic factors; Z i represents the value of the i-th neuron output by the LSTM neural network for the water injection volume of adjacent wells; K i represents the i-th convolutional kernel, * represents convolution, c represents the number of channels, and Ψ() is the response function of the fully connected layer, generally taken as the rectified linear unit (ReLu); The overall training of the model uses the Adaptive Moment Estimation (Adam) algorithm.

[0078] To improve the generalization ability of the model, different from the conventional model, the samples in model construction are not limited to a single well, but all single well samples after spatio-temporal fusion are substituted into the above model for modeling. The data samples are divided into training and testing according to a certain proportion. The training algorithm uses the Adaptive Moment Estimation (Adam) algorithm, and the loss function uses the mean squared error (mse) and the coefficient of determination (r 2 ). By selecting an appropriate batch size and the number of training times, the optimal learning model can be obtained.

[0079] Step 6: Model prediction

[0080] In the prediction method of the present invention, the influence of the change in the monthly working system on the development effect is considered among the influencing factors. During prediction, if the change in the working system can be given artificially, this change can be directly input into the model as an influencing factor. If there is no change in the working system, multi-step prediction for the future is carried out by the method of rolling completion.

[0081] Among them, since there is no change in the static factors, they are directly completed. For dynamic factors, such as stroke and pumping speed, etc., the moving average method is used for completion and used as the input of the model.

[0082] The following are several specific embodiments of applying the present invention

[0083] Embodiment 1

[0084] In a specific Embodiment 1 of applying the present invention, a prediction method for single-well development indexes of waterflooding oilfields integrating spatio-temporal information is established by using statistical data processing and machine learning methods for oilfield development index prediction.

[0085] As Figure 1 shown, it is the flow chart of the present invention for predicting single-well development indexes of oilfields by integrating spatio-temporal information.

[0086] In step 101, historical development data of 45 oil and water wells in a certain unit of GD Oilfield are selected to construct a sample library, and the importance of influencing factors is evaluated by the embedded method. Through three feature selection methods of decision tree, random forest, and Xgboost, features with a high degree of correlation and great influence on the prediction index are selected, the importance ranking of the input influencing factors of the model is determined, and the feature importance rankings of the three embedded methods as shown Figure 2 are obtained.

[0087] In step 102, the dynamic and static influencing factors are screened and determined by the voting method. Voting is carried out according to the ranking results of the previous embedded method, and the top 15 influencing factors affecting the development index are: converted mid-depth, formation pressure, formation temperature, permeability, remaining oil saturation, geological reserves, effective thickness, production days, pump depth, pump efficiency, displacement, stroke, pumping frequency, casing pressure, flowing fluid level;

[0088] In step 103, the spatial dynamic influencing factors are determined based on the serial correlation analysis. The relevant data of injection wells and production wells are extracted. Through the serial correlation analysis, the positive correlation influence of the production well and injection well sequences is obvious. As Figure 3 shows the correlation between the oil production and water injection of production well GO7-24C215 and injection well GO7-22-215 from May 2005 to April 2013. The correlation coefficient between the oil production and water injection at the 6th lag order is 0.67, and the influence is obvious.

[0089] In step 104, a water injection composite sample considering the lag of effectiveness is constructed. As Figure 4 shown, the absolute value m of the maximum lag order in the adjacent well set is selected; for example, for production well GO7-24C215 and injection well GO7-22-215, m is 6, so it will be lagged by 6 months. For all production wells, based on the latest time node of the training data, start to intercept the production oil volume data of the target production well for the previous 60 months as the output, and the corresponding other dynamic factor data as the input; for each adjacent injection well i corresponding to the production well, move forward by the corresponding number of months according to the absolute value of its maximum lag order, and intercept the monthly water injection volume of the injection well for 60 months; for production wells where the injection wells in the vicinity do not meet the threshold conditions (the situation where the water injection effectiveness cannot be clearly defined), the average injection volume sequence of its corresponding several injection wells is inserted into the input factors.

[0090] In step 105, a single-well development index prediction model that integrates spatio-temporal dynamic and static information is constructed based on LSTM and MLP, and machine learning models are used for experiments. The data samples are divided into training and testing at a ratio of 3:1. The training algorithm uses the Adaptive Moment Estimation (Adam) algorithm, and the loss function uses the mean square error (mse) and the coefficient of determination (r 2) The batch size is set to 50, the maximum number of complete training epochs is 300, and dropout is set to 0.1.

[0091] The mse of the fitting process is 0.0023, and r 2 is 0.99; the mse of the prediction process is 0.022, and r 2 is 0.82. The r 2 of single-well fitting is 0.99, and the r 2 of single-well prediction is 0.90.

[0092] The fitting process, fitting comparison graph, and prediction comparison graph of the model are as Figure 5 , Figure 6 , Figure 7 . Prediction accuracy: mse is 0.022, r 2 is 0.83. The prediction accuracy is 0.467 higher than that of the model considering only static information and dynamic time-series features (r 2 is 0.363). From the experimental results, it can be seen that the fitting and prediction accuracies of the machine learning model after fusing spatio-temporal information are greatly improved compared with the basic LSTM model, indicating the feasibility and advancement of this method.

[0093] In step 106, model prediction methods such as Figure 8 , Figure 9 , Figure 10 are used. Using the 60-month data from November 2014 to October 2019 as known data, a rolling average is performed forward to obtain the 12-month data from November 2019 to October 2020. The model is used for prediction, and the fitting, prediction, and unknown production prediction graphs are obtained. The mse of the fitting during the training process is 0.008, and the mse of the prediction is 0.006.

[0094] Example 2

[0095] In the specific implementation example 2 of applying the present invention, a prediction method for single-well development indexes of waterflooding oilfields that fuses spatio-temporal information is established by using statistical data processing and machine learning methods to predict oilfield development indexes.

[0096] As Figure 1 shown, it is the flow chart of predicting single-well development indexes of oilfields by fusing spatio-temporal information in the present invention.

[0097] In step 101, historical development data of 45 oil and water wells in a certain unit of the GD oilfield is selected to construct a sample library, and the importance of influencing factors is evaluated by the embedded method. Through three feature selection methods of decision tree, random forest, and Xgboost, features with a high degree of correlation and great influence on the prediction index are selected, and the importance ranking of the input influencing factors of the model is determined, obtaining as Figure 2Feature importance ranking of the three embedded methods shown

[0098] In step 102, the voting method is used to screen and determine dynamic and static influencing factors. Voting is carried out according to the sorting results of the previous embedded methods, and the top 15 influencing factors affecting the development index are obtained: converted middle depth, formation pressure, formation temperature, permeability, remaining oil saturation, geological reserves, effective thickness, production days, pump depth, pump efficiency, displacement, stroke, pumping frequency, casing pressure, flowing fluid level

[0099] In step 103, the spatial dynamic influencing factors are determined based on sequence correlation analysis. The relevant data of injection wells and production wells are extracted. Through sequence correlation analysis, the positive correlation influence of the oil-water well sequence is obvious. As Figure 12 shows the correlation between the oil production and water injection of production well GO7-34-186 and injection well GO7-35-166 from July 2019 to October 2020. The correlation coefficient between the oil production and water injection at the 7th lag order is 0.9182, and the influence is obvious

[0100] In step 104, a water injection composite sample considering the lag of effectiveness is constructed. As Figure 4 shown, the absolute value m of the maximum lag order in the adjacent well set is selected; for example, for production well GO7-34-186 and injection well GO7-35-166, m is 7, so it will lag for 7 months. For all production wells, based on the latest time node of the training data, start intercepting the production oil volume data of the target production well for the previous 60 months as the output, and the corresponding other dynamic factor data as the input; for each adjacent injection well i corresponding to the production well, move forward the corresponding number of months according to the absolute value of its maximum lag order, and intercept the monthly water injection volume of the injection well for 60 months; for production wells where the injection wells in the vicinity do not meet the threshold conditions (the situation where the water injection effectiveness cannot be clearly defined), insert the average value sequence of the injection volumes of its corresponding several injection wells into the input factors

[0101] In step 105, a single-well development index prediction model that combines spatio-temporal dynamic and static information is constructed based on LSTM and MLP, and machine learning models are used for experiments. The data samples are divided into training and testing at a ratio of 3:1. The training algorithm uses the Adaptive Moment Estimation (Adam) algorithm, the loss function uses the mean squared error (mse) and the coefficient of determination (r 2 ), the batch size is taken as 50, the maximum number of complete training times (epoch) is 300, and dropout is taken as 0.1

[0102] The mse in the fitting process is 0.0023, r 2 is 0.99; the mse in the prediction process is 0.022, r 2 is 0.82. The r 2is 0.99, and the r predicted by a single well is 2 is 0.90.

[0103] The fitting process, fitting comparison graph, and prediction comparison graph of the model are as Figure 5 , Figure 13 , Figure 14 . Prediction accuracy: mse is 0.020, r 2 is 0.85. The prediction accuracy is 0.430 higher than that of the model that only considers static information and dynamic time series features (r 2 is 0.420). From the experimental results, it can be seen that the fitting and prediction accuracy of the machine learning model after fusing spatio-temporal information are greatly improved compared with the basic LSTM model, indicating the feasibility and advancement of this method.

[0104] In step 106, the model prediction method is used, as Figure 15 shown. Taking the 60-month data from December 2006 to November 2011 as known data, performing a rolling average forward to obtain the 12-month data from December 2011 to November 2012, using the model for prediction, the obtained fitting, prediction, and unknown production prediction graphs, the mse of the fitting during the training process is 0.010, and the mse of the prediction is 0.008.

[0105] Example 3

[0106] In the specific Example 3 of applying the present invention, by using statistical data processing and machine learning methods, a prediction method for single-well development indexes of waterflooding oilfields that fuses spatio-temporal information is established to predict oilfield development indexes.

[0107] As Figure 1 shown, it is the flow chart of predicting single-well development indexes of oilfields by fusing spatio-temporal information in the present invention.

[0108] In step 101, historical development data of 45 oil and water wells in a certain unit of the GD oilfield are selected to construct a sample library, and the importance of influencing factors is evaluated by the embedded method. Through three feature selection methods of decision tree, random forest, and Xgboost, features with a high degree of correlation and great influence with the prediction index are selected, and the importance ranking of the input influencing factors of the model is determined, obtaining the feature importance ranking of the three embedded methods as Figure 2 shown.

[0109] In step 102, the voting method is used to screen and determine dynamic and static influencing factors. Voting according to the sorting results of the previous embedded method, the top 15 influencing factors affecting the development index are obtained: converted middle depth, formation pressure, formation temperature, permeability, remaining oil saturation, geological reserves, effective thickness, production days, pump depth, pump efficiency, displacement, stroke, pumping frequency, casing pressure, flowing fluid level;

[0110] In step 103, spatial dynamic influencing factors are determined based on sequence correlation analysis. The relevant data of injection wells and production wells are extracted. Through sequence correlation analysis, the positive correlation between the sequences of oil and water wells is significantly affected. For example, Figure 16 shows the correlation between the oil production and water injection of production well GO7-34N1306 and injection well GO7-33-306 from July 2012 to October 2020. The correlation coefficient between the oil production and water injection is 0.6688 at the 41st lag order, indicating a significant influence.

[0111] In step 104, a water injection composite sample considering the lag effect is constructed. As Figure 4 shown, the absolute value m of the maximum lag order in the set of adjacent wells is selected; for example, for production well GO7-34N1306 and injection well GO7-33-306, m is 41, so it needs to lag 41 months. For all production wells, based on the latest time node of the training data, start intercepting the oil production data of the target production well for the previous 60 months as the output, and the corresponding other dynamic factor data as the input; for each adjacent injection well i corresponding to a production well, move forward the corresponding number of months according to the absolute value of its maximum lag order, and intercept the monthly water injection volume of the injection well for 60 months; for production wells where the injection wells in the vicinity do not meet the threshold conditions (the situation where the water injection effect cannot be clearly determined), the average injection volume sequence of the corresponding several water wells is inserted into the input factors.

[0112] In step 105, a single-well development index prediction model that integrates spatio-temporal dynamic and static information is constructed based on LSTM and MLP, and machine learning models are used for experiments. The data samples are divided into training and testing at a ratio of 3:1. The training algorithm uses the Adaptive Moment Estimation (Adam) algorithm, the loss function uses the mean squared error (mse) and the coefficient of determination (r 2 ), the batch size is taken as 50, the maximum number of complete training epochs is 300, and dropout is taken as 0.1.

[0113] The mse in the fitting process is 0.0023, and r 2 is 0.99; the mse in the prediction process is 0.022, and r 2 is 0.82. The r 2 for single-well fitting is 0.99, and the r 2 for single-well prediction is 0.90.

[0114] The fitting process, fitting comparison chart, and prediction comparison chart of the model are as shown in Figure 5 , Figure 17 , Figure 18 . The prediction accuracy: mse is 0.014, and r 2 is 0.907. The prediction accuracy is higher than that of the model considering only static information and dynamic time series features (r 2is 0.500) and the height is 0.407. From the experimental results, it can be seen that the fitting and prediction accuracy of the machine learning model after fusing spatio-temporal information has been greatly improved compared with the basic LSTM model, indicating the feasibility and advancement of this method.

[0115] In step 106, model prediction methods such as Figure 19 are shown. Using the 60-month data from July 2011 to June 2016 as known data, perform a rolling average forward to obtain the 12-month data from July 2016 to June 2017, and use the model for prediction. The fitting, prediction, and unknown yield prediction graphs are obtained. The mse for fitting during the training process is 0.013, and the mse for prediction is 0.011.

[0116] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0117] Except for the technical features described in the specification, they are all known technologies to those skilled in the art.

Claims

1. A method for predicting the single - well production of a water - flooded oilfield by integrating spatio - temporal information, characterized in that, The single - well production prediction method for water - flooded oilfields integrating spatio - temporal information includes: Step 1: Evaluate the importance of influencing factors using the embedded method; Step 2: Screen and determine dynamic and static influencing factors using the voting method; Step 3: Determine spatial dynamic influencing factors based on sequence correlation analysis; Step 4: Construct a water - injection composite sample considering the lag effect; Step 5: Construct a single - well development index prediction model integrating spatio - temporal dynamic and static information based on LSTM and MLP; Step 6: Conduct model prediction.

2. The single-well production prediction method for waterflooding oilfields integrating spatio-temporal information according to claim 1, wherein In Step 1, for the dynamic and static influencing factors that affect the single - well production effect, the importance of the influencing factors is evaluated using the embedded method.

3. The single-well production prediction method for waterflooding oilfields integrating spatio-temporal information according to claim 2, wherein In Step 1, three algorithms based on tree models, namely decision tree, random forest, and Xgboost methods, are used to obtain the importance ranking of each influencing factor respectively.

4. The method for predicting the single - well production of a water - flooded oilfield integrating spatio - temporal information according to claim 1, wherein, In Step 2, the weighted voting method is used for screening. The principle of the weighted voting method is to assign a weight reflecting its overall performance to each individual learner, and its overall performance is measured according to the validation set. In this way, the votes of more accurate models will be effectively counted multiple times.

5. The single-well production prediction method for waterflooding oilfields integrating spatio-temporal information according to claim 4, characterized in that, In Step 2, the formula of the weighted voting method is expressed as follows: Among them, where: x is the input test sample; ω i is the weight of the classifier c i , which is one third in the present invention; represents the probability that the base classifier i predicts that x belongs to the class j; j * is the maximum probability that the voting classifier finally predicts that x belongs to the class j.

6. The method for predicting the single-well production of a waterflooding oilfield integrating spatio-temporal information according to claim 1, characterized in that, In Step 3, the time series of water injection volume that is significantly correlated, including spatial and dynamic factors, is regarded as the factor affecting the well dynamics; due to the existence of the spatial well spacing between oil wells and water wells, the dynamics of water wells are reflected in oil wells with a certain time lag.

7. The single-well production prediction method for waterflooding oilfields integrating spatio-temporal information according to claim 6, characterized in that In Step 3, for the oil well whose production is to be predicted, according to the geodetic coordinates of the oil well, the well numbers of water wells within a 500 - meter radius of the oil well are screened through an algorithm in the database to obtain the corresponding relationship table between the oil well and its adjacent water wells; at this time, one oil well may correspond to multiple water wells.

8. The method for predicting the single - well production of a water - flooded oilfield integrating spatio - temporal information according to claim 7, wherein, In Step 3, in the table of the oil well and the corresponding water wells, the time series of the liquid production volume of this oil well and the water injection volume of all corresponding water wells are extracted according to the well numbers. The sequence correlation analysis with different lag times is carried out for each water well to determine the correlation coefficient of different lag - time sequences of each water well.

9. The single-well production prediction method for waterflooding oilfields integrating spatio-temporal information according to claim 8, characterized in that, In Step 3, select the water injection wells with a lag order < 12 orders, the absolute value of the correlation coefficient is the largest and greater than 0.6 as the corresponding water wells of the oil well; the lag order at which the absolute value of the correlation coefficient is the largest is m; for the oil wells where the injection wells in the vicinity do not meet the threshold conditions, that is, the situation where the water injection effect cannot be clearly determined, the average value sequence of the injection volumes of these water wells is inserted into the input factors to unify the input sequence format.

10. The water drive oilfield single well production prediction method integrating spatio-temporal information according to claim 1, characterized in that In Step 4, for all oil wells, taking the latest time node of the training data as the reference point, start intercepting the oil production data of the target oil well for 60 months backward as the output, and the corresponding other dynamic factor data as the input; For each adjacent injection well i corresponding to an oil well, according to the lag order m when the correlation coefficient is the largest that has been determined, start intercepting the water injection volume data of the injection well for 60 months m months backward based on the time period corresponding to the oil well; analyze all the oil wells and water wells in a block or an oilfield to construct a composite sample.

11. The method for predicting the single well production of a water drive oilfield integrating spatio-temporal information according to claim 1, wherein, In step 5, based on the multi-layer perceptron network MLP and the LSTM network, a machine learning network model for dynamic and static information fusion is constructed; the neuron values after being processed by the MLP network and the LSTM network are weighted and integrated to obtain the output formula of the machine learning model as follows: Among them, X i is the value of the i-th output neuron processed by the MLP neural network for static features; Y i represents the value of the i-th neuron output by the LSTM neural network for dynamic factors; Z i represents the value of the i-th neuron output by the LSTM neural network for the water injection volume of adjacent wells; K i represents the i-th convolution kernel, * represents convolution, c represents the number of channels, Ψ() is the response function of the fully connected layer, taken as the rectified linear unit ReLu; the overall training of the model uses the adaptive moment estimation Adam algorithm.

12. The method for predicting the single-well production of a waterflooding oilfield integrating spatio-temporal information according to claim 11, wherein, In step 5, the samples for model construction are not limited to a single well, but all the single well samples after spatio-temporal fusion are substituted into the above model for modeling; the data samples are divided into training and testing according to a certain proportion, the training algorithm uses the Adaptive Moment Estimation Adam algorithm, the loss functions use the mean squared error mse and the coefficient of determination r2, and by selecting an appropriate batch size and the number of training times, the optimal learning model can be obtained.

13. The water flooding oilfield single well production prediction method integrating spatio-temporal information according to claim 1, wherein In step 6, the influence of the change in the monthly working system on the development effect is considered among the influencing factors; during prediction, if the change in the working system can be given artificially, this change can be directly input into the model as an influencing factor; if there is no change in the working system, the future multi-step prediction is carried out by the method of rolling completion.

14. The method for predicting the single-well production of a waterflooding oilfield integrating spatio-temporal information according to claim 13, wherein In step 6, during completion, the static factors are directly completed because there is no change; the dynamic factors include the stroke and the pumping frequency, and the moving average method is used for completion and used as the input of the model.

15. A single-well production prediction system for waterflooding oilfields that integrates spatio-temporal information, characterized in that, The water flooding oilfield single well production prediction system integrating spatio-temporal information uses the water flooding oilfield single well production prediction method described in any one of claims 1-14 to predict the water flooding oilfield single well production.

Images