A data-driven coalbed methane production prediction method and system

Abstract

The present application relates to a kind of coalbed gas production prediction method and system based on data driving, first, the coalbed gas dynamic production history data of the predicted production well and the coalbed gas production prediction related data of the predicted production well are acquired;According to the coalbed gas dynamic production history data, the production well type to which the predicted production well belongs is determined;Then select the trained coalbed gas production prediction model corresponding to the production well type as target model;Finally, the coalbed gas production prediction related data of the predicted production well is input into the target model, and the production data of the predicted production well is obtained.The present application obtains the most easily obtained and the most valuable monitoring data in coalbed gas mining site, calculates the coalbed gas production according to this part of data and using the production prediction model trained by machine learning method, reduces the data acquisition time, and improves the coalbed gas production prediction speed.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development, and in particular to a data-driven method and system for predicting coalbed methane production. Background Technology

[0002] Currently, the global energy consumption pattern is shifting from traditional fossil fuels like coal and oil to clean energy sources like natural gas. Coalbed methane (CBM), as a crucial component of clean energy, possesses abundant reserves and broad development prospects, holding strategic significance for global sustainable development and environmental protection. In the long-term development of unconventional oil and gas resources, accurate production forecasting is fundamental for assessing reservoir development, optimizing production measures, adjusting injection and production schemes, and formulating secondary development strategies, directly impacting the economics, efficiency, and sustainability of reservoir development. However, due to its unique adsorption-desorption characteristics and complex flow and transport mechanisms, CBM production forecasting presents significant challenges.

[0003] Traditional physics-driven methods primarily include decline curve analysis and numerical simulation. Decline curve analysis is ill-suited to complex geological structures and multi-field coupling, exhibiting overly idealized predictions. Geological modeling in numerical simulation relies on geological parameters and fluid data, which often require significant time and resources to acquire. Furthermore, numerical simulation techniques necessitate extensive parameter tuning and historical data fitting based on expert experience. Therefore, current traditional numerical simulation techniques and decline curve analysis methods cannot quickly and accurately predict coalbed methane reservoir production. Summary of the Invention

[0004] The purpose of this invention is to provide a data-driven method and system for predicting coalbed methane production. Based on the most easily obtained and valuable monitoring data from coalbed methane collection sites, a self-learning model is established through machine learning to automatically capture and learn the implicit features in the data, thereby predicting coalbed methane production more quickly and accurately.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A data-driven method for predicting coalbed methane production includes:

[0007] Acquire historical dynamic coalbed methane (CBM) production data and CBM production prediction data for the well to be predicted. The CBM production prediction data includes dynamic historical data and static data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes post-fracturing permeability, original gas content, geostress, and reservoir thickness.

[0008] Based on the historical dynamic production data of coalbed methane, the production well type to which the production well to be predicted belongs is determined; the production well types include: conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells; conventional production wells are those whose production successively goes through three stages: a production increase period, a stable production period, and a production decrease period; single-peak production wells are those whose stable production period is less than a set threshold or whose stable production period is zero; double-peak production wells are those that include two production increase periods and two production decrease periods, and whose stable production period is less than a set threshold or whose stable production period is zero; irregular production wells are those that do not belong to the conventional production wells, single-peak production wells, or double-peak production wells;

[0009] A pre-trained coalbed methane production prediction model corresponding to the type of production well is selected as the target model; the pre-trained coalbed methane production prediction model is a model trained with coalbed methane production prediction data of sample production wells as input and actual coalbed methane production data of sample production wells as labels; the pre-trained coalbed methane production prediction model adopts a bidirectional long short-term memory network model.

[0010] Input the coalbed methane production prediction data of the well to be predicted into the target model to obtain the production data of the well to be predicted; the production data includes bottom hole flowing pressure, gas production and water production.

[0011] Optionally, before selecting the trained coalbed methane production prediction model corresponding to the production well type as the target model, the method further includes training the coalbed methane production prediction model, the training process of which is as follows:

[0012] Obtain relevant data on coalbed methane production forecasts and actual coalbed methane production data of sample production wells; the sample production wells include conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells.

[0013] For each type of production well, the coalbed methane production prediction model is trained using the coalbed methane production prediction data of the sample production well as the actual coalbed methane production data of the sample production well as the label, thus obtaining the trained coalbed methane production prediction model corresponding to the production well type.

[0014] Optionally, the loss function for the training process is:

[0015] L tol =μLs+τLd

[0016] Among them, L tol Total loss; L d and L sLet μ and τ represent the dynamic loss and static loss, respectively; μ and τ represent the coefficients of the static loss and dynamic loss, respectively, and μ + τ = 1.

[0017] The dynamic loss is:

[0018]

[0019] Here, Net represents the computation process of a bidirectional long short-term memory network using the chain rule of differentiation and the backpropagation algorithm; Net Pw This refers to the bottom hole flowing pressure predicted based on a coalbed methane production prediction model; Net Qg This represents the gas production predicted based on a coalbed methane production prediction model; Net Qw R represents the water production predicted based on the coalbed methane production prediction model; N is the total sample size; R Pw R represents the actual bottom hole flowing pressure; Qg Represents the actual gas production value; R Qw This represents the actual water production value;

[0020] The static loss is:

[0021]

[0022] Where K is the permeability after fracturing; Gc is the original gas content; Gs is the geostress; and Rt is the reservoir thickness.

[0023] Optionally, before obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, the method further includes selecting sample production wells, and the selection method is as follows:

[0024] The development status of the selected sample production wells is evaluated using evaluation parameters to obtain evaluation results; the evaluation parameters include development time, production capacity, and percentage of effective time.

[0025] Select the candidate production wells whose evaluation results meet the preset conditions as the target production wells.

[0026] The above method can remove production wells with poor production conditions, thereby removing the corresponding low-quality data and improving the model's prediction accuracy.

[0027] Optionally, after obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, the method further includes: preprocessing the sample data; the sample data includes the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells.

[0028] Optionally, the preprocessing of the sample data includes:

[0029] Detect outliers in the sample data;

[0030] The outlier was removed from the sample data.

[0031] By detecting and removing outliers in sample data, the quality of the sample data can be improved, thereby increasing the prediction accuracy of the yield prediction model.

[0032] Optionally, the preprocessing of the sample data further includes filling in missing values ​​in the sample data:

[0033] Identify the missing data segments in the sample data; the missing data segments include gas production and water production.

[0034] Obtain the missing segment completion related data corresponding to the missing segment data; the missing segment completion related data is used to calculate the missing segment data; the missing segment completion related data includes bottom hole flowing pressure, oil pressure, tubing pressure, and dynamic fluid level;

[0035] The missing segment completion data is input into the trained completion model to obtain the numerical values ​​corresponding to the missing segment data. The trained completion model is a model trained using the bottom hole flowing pressure, oil pressure, tubing pressure and dynamic fluid level in the normal data as inputs and the gas production and water production in the normal data as labels.

[0036] This invention also provides a data-driven coalbed methane production prediction system, comprising:

[0037] The data acquisition module for the production well to be predicted is used to acquire historical dynamic coalbed methane production data and coalbed methane production prediction data of the production well to be predicted. The coalbed methane production prediction data includes dynamic historical data and static data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes permeability after fracturing, original gas content, geostress, and reservoir thickness.

[0038] The module for determining the type of production well to be predicted is used to determine the type of production well to which the production well to be predicted belongs based on the historical dynamic production data of coalbed methane. The production well types include: conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells. Conventional production wells are those whose production successively goes through three stages: a production increase period, a stable production period, and a production decrease period. Single-peak production wells are those whose stable production period is less than a set threshold or whose stable production period is zero. Double-peak production wells are those that include two production increase periods and two production decrease periods, and whose stable production period is less than a set threshold or whose stable production period is zero. Irregular production wells are those that do not belong to the conventional production wells, single-peak production wells, or double-peak production wells.

[0039] The target model selection module is used to select a pre-trained coalbed methane production prediction model corresponding to the type of production well as the target model; the pre-trained coalbed methane production prediction model is a model trained with coalbed methane production prediction data of sample production wells as input and actual coalbed methane production data of sample production wells as labels; the pre-trained coalbed methane production prediction model adopts a bidirectional long short-term memory network model.

[0040] The production data prediction module is used to input the coalbed methane production prediction data of the well to be predicted into the target model to obtain the production data of the well to be predicted; the production data includes bottom hole flowing pressure, gas production and water production.

[0041] Optionally, the system further includes: a training module, used to train the coalbed methane production prediction model before selecting the trained coalbed methane production prediction model corresponding to the production well type as the target model;

[0042] The training module specifically includes:

[0043] The sample production well data acquisition submodule is used to acquire coalbed methane production prediction data and actual coalbed methane production data of the sample production wells; the sample production wells include conventional production wells, single-peak production wells, double-peak production wells and irregular production wells.

[0044] The production prediction model training submodule is used to train the coalbed methane production prediction model from both positive and negative directions for each type of production well, using the coalbed methane production prediction data of the sample production well as the actual coalbed methane production data of the sample production well as the label, so as to obtain a trained coalbed methane production prediction model.

[0045] Optionally, the training module further includes:

[0046] The sample production well selection submodule is used to evaluate the development status of the candidate sample production wells using evaluation parameters before obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, and to obtain the evaluation results. The evaluation parameters include development time, production capacity and effective time percentage. The candidate sample production wells whose evaluation results meet the preset conditions are selected as target sample production wells.

[0047] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0048] This invention provides a data-driven method and system for predicting coalbed methane (CBM) production. First, it acquires historical dynamic CBM production data and relevant CBM production prediction data for the well to be predicted. Based on the historical dynamic CBM production data, it determines the type of well to which the well belongs; the well types include: conventional wells, single-peak wells, double-peak wells, and irregular wells. Then, it selects a pre-trained CBM production prediction model corresponding to the well type as the target model. Finally, it inputs the relevant CBM production prediction data of the well to be predicted into the target model to obtain the production data of the well. This invention obtains the most readily available and valuable monitoring data from CBM extraction sites, and uses this data, along with a production prediction model trained using machine learning methods, to calculate CBM production, thus reducing data acquisition time and improving the speed of CBM production prediction. Furthermore, this invention uses a pre-trained production prediction model for CBM production prediction, which can automatically capture and learn implicit features in the data, thereby improving the accuracy of CBM production prediction. Moreover, this invention provides a pre-trained production prediction model for each type of production well. When predicting the coalbed methane production of a production well, the production prediction model corresponding to the type of production well is selected for prediction, which further improves the accuracy of production prediction. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 The flowchart of the data-driven coalbed methane production prediction method provided in Embodiment 1 of the present invention is shown below.

[0051] Figure 2 This is a schematic diagram of four types of coalbed methane production curves provided in Embodiment 1 of the present invention;

[0052] Figure 3 This is a diagram illustrating the effect of dynamic data outlier detection provided in Embodiment 1 of the present invention.

[0053] Figure 4 This is a flowchart illustrating the construction process of the intelligent data completion model provided in Embodiment 1 of the present invention.

[0054] Figure 5 This is a schematic diagram of a data cube representing dynamic monitoring data of a coalbed methane reservoir provided in Embodiment 1 of the present invention;

[0055] Figure 6 The flowchart of the intelligent prediction method for coalbed methane production based on a bidirectional long short-term memory network, which takes into account dynamic and static losses, is provided in Embodiment 1 of the present invention.

[0056] Figure 7 This is a graph showing the predicted coalbed methane production rate provided in Embodiment 1 of the present invention.

[0057] Figure 8 The impact of data governance provided in Embodiment 1 of the present invention on the long-term and short-term forecasts of the production forecasting model. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The purpose of this invention is to provide a data-driven method and system for predicting coalbed methane production. Based on the most easily obtained and valuable monitoring data from coalbed methane collection sites, a self-learning model is established through machine learning to automatically capture and learn the implicit features in the data, thereby predicting coalbed methane production more quickly and accurately.

[0060] This invention addresses the limitations of traditional numerical simulation techniques and declining curve analysis in rapidly and accurately predicting coalbed methane (CBM) reservoir production. It innovatively proposes a data-driven CBM production prediction method that integrates data governance. This method encompasses a preliminary assessment of the development status of producing wells, the definition of a CBM production curve model, the establishment of a physics-guided dynamic data anomaly detection and intelligent completion method, and the development of an intelligent CBM production prediction algorithm based on a bidirectional long short-term memory network, considering both dynamic and static losses. The designed production prediction system can handle complex geological conditions and offers advantages such as fast prediction speed, high efficiency, and strong robustness. Furthermore, the established data governance system improves data quality, enhances short-term prediction accuracy, and increases long-term prediction robustness.

[0061] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] Example 1

[0063] This embodiment provides a data-driven method for predicting coalbed methane production. Please refer to [link to relevant documentation]. Figure 1 ,include:

[0064] S1. Obtain historical dynamic production data of coalbed methane (CBM) from the well to be predicted, as well as relevant data for CBM production prediction. The relevant data for CBM production prediction includes dynamic historical data and static data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes post-fracturing permeability, initial gas content, geostress, and reservoir thickness.

[0065] S2. Based on the historical dynamic production data of coalbed methane, determine the type of production well to which the production well to be predicted belongs. The production well types include: conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells; conventional production wells are those whose production successively goes through three stages: a production increase period, a stable production period, and a production decrease period; single-peak production wells are those whose stable production period is less than a set threshold or whose stable production period is zero; double-peak production wells are those that include two production increase periods and two production decrease periods, and whose stable production period is less than a set threshold or whose stable production period is zero; irregular production wells are those that do not belong to the conventional production wells, single-peak production wells, or double-peak production wells.

[0066] S3. Select the pre-trained coalbed methane production prediction model corresponding to the type of production well as the target model. The pre-trained coalbed methane production prediction model is a model trained using coalbed methane production prediction data of the sample production wells as input and actual coalbed methane production data of the sample production wells as labels; the pre-trained coalbed methane production prediction model adopts a bidirectional long short-term memory network model.

[0067] S4. Input the coalbed methane production prediction data of the well to be predicted into the target model to obtain the production data of the well to be predicted. The production data includes bottom hole flowing pressure, gas production, and water production.

[0068] As an optional implementation, before selecting the trained coalbed methane production prediction model corresponding to the production well type as the target model, the method further includes: S0, training the coalbed methane production prediction model, the training process being as follows:

[0069] S01. Obtain relevant data on coalbed methane production prediction of sample production wells and actual coalbed methane production data of sample production wells; the sample production wells include conventional production wells, single-peak production wells, double-peak production wells and irregular production wells.

[0070] In this embodiment, a coalbed methane production prediction model needs to be trained for each type of production well. Therefore, the selected sample production wells must include each type of production well.

[0071] S02. For each type of production well, using the coalbed methane production prediction data of the sample production well as the input and the actual coalbed methane production data of the sample production well as the label, train the coalbed methane production prediction model to obtain the trained coalbed methane production prediction model corresponding to the production well type.

[0072] As an optional implementation, before acquiring the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, the method further includes selecting sample production wells, and the selection method is as follows:

[0073] The development status of the selected sample production wells is evaluated using evaluation parameters to obtain evaluation results; the evaluation parameters include development time, production capacity, and percentage of effective time.

[0074] Select the candidate production wells whose evaluation results meet the preset conditions as target production wells.

[0075] It should be noted that before training the coalbed methane production prediction model, sample data needs to be selected first. The sample data is obtained based on sample production wells. Therefore, in this embodiment, some production wells with good production conditions are first obtained as target production wells, and then sample data of the target production wells are obtained for model training.

[0076] Because the production status of various wells in a coalbed methane reservoir varies greatly, wells with poor production status receive less attention from the oilfield, and the data they provide is of poor quality. Therefore, it is necessary to establish a development status assessment method based on expert experience. This embodiment assesses the development status of a well from three aspects: development time (T), production capacity (PA), and effective time percentage (ER). For any production well, the following equation must be satisfied:

[0077]

[0078] Where SI represents the well status (0 for well shut-in, 1 for well open), q gThe daily gas production is defined as the production volume. Based on the above constraints, wells with poor production conditions were screened out, while wells with good production capacity were retained. A total of 1520 wells were collected in this experiment, and 798 producing wells that met the conditions were selected using the above method.

[0079] It should be noted that the above-mentioned effective time percentage of 0.2 or greater is an optimal value calculated through experiments. 0.2 can be replaced with other values ​​depending on the actual situation. The same applies to 90 days and a production capacity of 100 cubic meters. This embodiment does not limit the specific values.

[0080] After obtaining the target sample production wells, the target sample production wells are divided into four types. The data of each type of sample production well is used to train the coalbed methane production prediction model corresponding to that type.

[0081] Coalbed methane production curve patterns can be broadly classified into four types: I, II, III, and IV (see details). Figure 2 Each type of production curve corresponds to one of the aforementioned production wells. Type I represents the most common production pattern of coalbed methane, experiencing three stages: increased production, stable production, and decreased production. Unlike Type I, Type II curves exhibit a very short, or even nonexistent, stable production period. This is mainly due to low reservoir permeability and inadequate drainage schemes leading to weak gas supply capacity. Type III curves are characterized by two peaks. This is primarily due to the high saturation of free gas in the coalbed methane reservoir fractures, which is discharged along with water during the early drainage stage, causing a rapid increase in production. After the free gas is discharged, production decreases, forming the first peak. Subsequently, adsorbed gas in the coalbed methane reservoir begins to desorb, causing production to increase again, forming the second peak. Type IV curves represent an irregular production pattern. This is because the interaction of drainage systems, reservoir geology, and fracturing measures during reservoir development results in a complex and irregular production pattern. Compared to the other three curve types, Type IV curves generally have relatively lower production. In this experiment, dynamic data from a total of 782 wells were collected, with 252, 222, 159 and 149 curves classified as Class I, II, III and IV, respectively.

[0082] After obtaining the target sample production wells, it is necessary to acquire data for each target sample production well, i.e., sample data. Sample data includes coalbed methane (CBM) production prediction data and actual CBM production data for the sample production wells. Specifically, the CBM production prediction data for the sample production wells includes both dynamic and static historical data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes post-fracture permeability, initial gas content, geostress, and reservoir thickness.

[0083] It should be noted that the sample data includes input data for the production prediction model and labeled data for model training. The dynamic historical data used as model input can be the bottomhole flowing pressure, gas production, water production, oil pressure, tubing pressure, and dynamic fluid level for each day of the previous month for the sample well, as well as the cumulative oil production and cumulative water production for the previous month. The model's labels are the bottomhole flowing pressure, gas production, and water production for the month following the month corresponding to the input data. The principle is to predict future coalbed methane production using static data and the dynamic historical data from previous sample wells. The static data for each producing well remains unchanged.

[0084] After obtaining the sample data, further processing is required to improve the prediction accuracy of the yield prediction model.

[0085] As an optional implementation, after obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, the method further includes: preprocessing the sample data; the sample data includes the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells.

[0086] In some embodiments, the preprocessing of the sample data includes:

[0087] Detect outliers in the sample data;

[0088] The outlier was removed from the sample data.

[0089] The specific methods for detecting and removing outliers in sample data can be as follows:

[0090] A three-dimensional coordinate system is established using gas production (Qg), water production (Qw), and bottom hole flowing pressure (Pw). The dynamic data of a single well is then mapped onto this three-dimensional coordinate system based on the number of days the well has been in operation. First, the distance between any two points (p, o) is calculated using the following formula:

[0091]

[0092] The distance between point o and its k-th nearest point is defined as d. k All conditions satisfying the condition that the distance from p to o does not exceed d k The set of points is defined as N k (o), the formula is:

[0093] N k (o)={p∈Z|d(p[,o)≤d k}

[0094] Where Z represents the set of all mapped points of a single well in the three-dimensional coordinate system. In this experiment, the fluctuation range of outliers was considered based on expert experience, and k = 15. Then, the absolute distance from each point p to o was determined, and the formula for the absolute distance is:

[0095] Ad k (p, o) = max{d k ,d(p,o)}

[0096] The density (I) and anomaly (Abd) of this point can be expressed as:

[0097]

[0098] The outlier of all points is calculated using the method described above and arranged in descending order. A higher outlier indicates a higher probability that the point is an outlier. Expert experience is used to determine the contamination level (ct) of the production well data, which is the proportion of contaminated data collected from that well. Therefore, the amount of contaminated data can be determined by ct*T. Contaminated data is then filtered out based on the outlier, and the filtered outliers are deleted. Figure 3 As shown in the figure, the left figure shows obvious outliers in the actual gas production curve, while the right figure shows the results of outlier detection based on the above method.

[0099] As an optional implementation, the preprocessing of the sample data further includes imputing missing values ​​in the sample data:

[0100] Identify the missing data segments in the sample data; the missing data segments include gas production and water production.

[0101] Obtain the missing segment completion related data corresponding to the missing segment data; the missing segment completion related data is used to calculate the missing segment data; the missing segment completion related data includes bottom hole flowing pressure, oil pressure, tubing pressure and dynamic fluid level.

[0102] The missing segment completion data is input into the trained completion model to obtain the numerical values ​​corresponding to the missing segment data. The trained completion model is a model trained using the bottom hole flowing pressure, oil pressure, tubing pressure and dynamic fluid level in the normal data as inputs and the gas production and water production in the normal data as labels.

[0103] Outlier detection can identify outliers in the dynamic data of each production well. In addition, the production curves of each well may contain data gaps due to human or environmental factors, such as... Figure 4 The diagram illustrates the construction of an intelligent data completion model. Its core idea is to utilize the relationships between various variables in the normal data of a single well, and construct an intelligent completion model using an extreme learning machine (XGboost). The dynamic parameters input to the model include bottomhole flowing pressure (Pw), oil pressure (Po), tubing pressure (Pc), and dynamic fluid level (D). l The model outputs gas production (Qg) and water production (Qw). An intelligent learning model is trained based on the relationships between variables in normal data, and the optimal model is obtained after training. Then, based on missing and outlier data, the bottom hole flowing pressure (Pw), oil pressure (Po), tubing pressure (Pc), and dynamic fluid level (D) are analyzed. l Using ) as input, the gas production (Qg) and water production (Qw) are predicted through intelligent completion, and then the missing gas production and water production in a single well can be completed.

[0104] To reduce computational costs, this embodiment can also generate a data cube based on the dynamic monitoring data of the coalbed methane reservoir before training the production prediction model. For example... Figure 5 As shown, for each production well in a coalbed methane reservoir, its dynamic monitoring data is a two-dimensional array. The horizontal axis (x1 to x8) represents the dynamic parameters of the well, including bottom hole flowing pressure (Pw), gas production (Qg), water production (Qw), oil pressure (Po), tubing pressure (Pc), and dynamic fluid level (D). lThe dataset contains eight variables: cumulative oil production (Co), cumulative water production (Cw), and cumulative water production (Cw). The vertical axis represents the development time (T). Each producing well generates a corresponding sample dataset, with each sample generated by traversing the dataset using two data extraction windows (input window W_I and output window W_O). The input window (W_I) is w1×8, where w1 is the time length of the input window and 8 represents the number of the aforementioned feature parameters. The output window (W_O) is w2×3, where w2 is the time length of the output window and 3 represents the number of output variables to be predicted, namely bottomhole flowing pressure (Pw), gas production (Qg), and water production (Qw). The input and output windows traverse the entire dataset along the vertical time axis with a step size of 1, thus forming the dynamic input set (Xd) and output set (Y) for the corresponding producing well. In addition to the dynamic monitoring data mentioned above, this invention also constructs a static input set (Xs) for each production well. The parameters in the static dataset include the permeability after fracturing (K), the original gas content (Gc), the geostress (Gs), and the reservoir thickness (Rt). In this experiment, the dataset based on all 782 production wells is D (D is defined as D = [Xd + Xs] + Y). Based on different curve patterns, four datasets were constructed, namely D1, D2, D3, and D4. It is worth noting that each of these datasets consists of an input set (dynamic input set Xd and static input set Xs) and an output set (Y).

[0105] The coalbed methane production prediction model in this embodiment adopts a prediction model based on a bidirectional long short-term memory network that considers both dynamic and static losses. For example... Figure 6 As shown, the coalbed methane production prediction model can be divided into four parts. The first part is the input, which consists of dynamic correlation historical data and static correlation data. The dynamic correlation historical data includes bottom hole flowing pressure (Pw), gas production (Qg), water production (Qw), oil pressure (Po), tubing pressure (Pc), and dynamic fluid level (D). l The first part contains the cumulative oil production (Co) and cumulative water production (Cw); static correlation data includes post-fracturing permeability (K), initial gas content (Gc), geostress (Gs), and reservoir thickness (Rt). The second part consists of a bidirectional long short-term memory network (Bi-LSTM), which includes forward and backward propagation. Both the forward and backward propagation layers are composed of 32 long short-term memory units (LSTM), with σ being the sigmoid activation function. The third part is the output, constructed from three neurons, representing the predicted bottomhole flowing pressure (Pw), gas production (Qg), and water production (Qw), respectively. The final part is the physical guidance layer, which includes dynamic and static losses.

[0106] Optionally, the loss function for the production prediction model training process is:

[0107] L tol =μLs+τLd

[0108] Among them, L tol Total loss; L d and L s Let μ and τ represent the dynamic loss and static loss, respectively; μ and τ represent the coefficients of the static loss and dynamic loss, respectively, and μ + τ = 1. In this experiment, the model achieved the highest accuracy when the values ​​of μ and τ were 0.72 and 0.28, respectively, after multiple trials.

[0109] The dynamic loss is:

[0110]

[0111] Here, Net represents the computation process of a bidirectional long short-term memory network using the chain rule of differentiation and the backpropagation algorithm; Net Pw This refers to the bottom hole flowing pressure predicted based on a coalbed methane production prediction model; Net Qg This represents the gas production predicted based on a coalbed methane production prediction model; Net Qw R represents the water production predicted based on a coalbed methane production prediction model. Pw R represents the actual bottom hole flowing pressure; Qg Represents the actual gas production value; R Qw This represents the actual water production value; N is the total number of samples; it is worth noting that the total number of samples N is different for different datasets, such as D, D1, D2, D3 and D4.

[0112] The static loss is:

[0113]

[0114] Where K is the permeability after fracturing; Gc is the original gas content; Gs is the geostress; and Rt is the reservoir thickness.

[0115] This embodiment constructs four datasets: D1, D2, D3, and D4, representing four curve types. Within each dataset, the dataset is divided into a training set and a test set in a 9:1 ratio based on the total number of samples; that is, the training set comprises 9 / 10 of the total samples, and the test set comprises 1 / 10. The core of the intelligent coalbed methane production prediction model developed based on Bi-LSTM includes a weight matrix W and a threshold matrix B. First, W and B are initialized, and the input data from the training set is imported into the model. Based on the weight matrix W and the threshold matrix B, the model's prediction results can be obtained, including bottomhole flowing pressure (Pw), gas production (Qg), and water production (Qw). The model is then used to calculate the predicted values ​​based on a custom loss function L. tolThe model error can be calculated with the current weight matrix W and threshold matrix B. Then, W and B are updated based on the error value using the backpropagation algorithm, gradually reducing the loss function. This process is repeated until the loss function stabilizes and stops decreasing, at which point model training terminates. Finally, the accuracy of the trained model is evaluated using a test set, using the coefficient of determination (R²). 2 Its robustness is verified using three statistical indicators: mean absolute error (MAE), root mean square error (RMSE), and mean square error. Figure 7 As shown, it is evident that the production prediction results (Bi-LSTM) based on the production prediction method of this invention have a high degree of fit with the actual gas production, demonstrating satisfactory prediction accuracy, and exhibiting better prediction performance compared to the conventional Long Short-Term Memory (LSTM) method. Figure 8 The impact of data governance on coalbed methane production prediction models is presented. It can be seen that for short-term predictions (1-5 months), the data after governance can improve the prediction accuracy of the model. More importantly, for long-term predictions, the data after governance can significantly enhance the robustness of the model.

[0116] The beneficial technical effects of this invention are as follows:

[0117] 1. This invention designs a data-driven coalbed methane production prediction framework that integrates data governance, which can accurately and efficiently predict the production of coalbed methane reservoirs.

[0118] 2. This invention proposes a method for assessing the development status of production wells, which can quickly and effectively assess the development status of production wells in coalbed methane reservoirs and is conducive to improving data quality.

[0119] 3. This invention defines four types of production curve patterns for coalbed methane reservoir production wells. The curve patterns are divided into four categories according to different reservoir geological conditions and mining measures, which can effectively improve the prediction accuracy.

[0120] 4. This invention establishes a complete physical guidance-based coalbed methane reservoir data governance system, which can automatically and intelligently detect outliers and complete missing values ​​in dynamic data, improve data quality, and enhance the performance of the model in short-term and long-term prediction.

[0121] 5. This invention develops an intelligent prediction method for coalbed methane production based on a bidirectional long short-term memory network that takes into account dynamic and static losses. It combines dynamic and static losses to guide the training of the bidirectional long short-term memory network, and the bidirectional transmission mechanism in the network can better capture data features and improve prediction accuracy.

[0122] Example 2

[0123] This embodiment provides a data-driven coalbed methane production prediction system, including:

[0124] The data acquisition module for the production well to be predicted is used to acquire historical dynamic coalbed methane production data and coalbed methane production prediction data of the production well to be predicted. The coalbed methane production prediction data includes dynamic historical data and static data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes permeability after fracturing, original gas content, geostress, and reservoir thickness.

[0125] The module for determining the type of production well to be predicted is used to determine the type of production well to which the production well to be predicted belongs based on the historical dynamic production data of coalbed methane. The production well types include: conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells. Conventional production wells are those whose production successively goes through three stages: a production increase period, a stable production period, and a production decrease period. Single-peak production wells are those whose stable production period is less than a set threshold or whose stable production period is zero. Double-peak production wells are those that include two production increase periods and two production decrease periods, and whose stable production period is less than a set threshold or whose stable production period is zero. Irregular production wells are those that do not belong to the conventional production wells, single-peak production wells, or double-peak production wells.

[0126] The target model selection module is used to select a pre-trained coalbed methane production prediction model corresponding to the type of production well as the target model; the pre-trained coalbed methane production prediction model is a model trained with coalbed methane production prediction data of sample production wells as input and actual coalbed methane production data of sample production wells as labels; the pre-trained coalbed methane production prediction model adopts a bidirectional long short-term memory network model.

[0127] The production data prediction module is used to input the coalbed methane production prediction data of the well to be predicted into the target model to obtain the production data of the well to be predicted; the production data includes bottom hole flowing pressure, gas production and water production.

[0128] As an optional implementation, the system further includes: a training module, used to train the coalbed methane production prediction model before selecting the trained coalbed methane production prediction model corresponding to the production well type as the target model;

[0129] The training module specifically includes:

[0130] The sample production well data acquisition submodule is used to acquire coalbed methane production prediction data and actual coalbed methane production data of the sample production wells; the sample production wells include conventional production wells, single-peak production wells, double-peak production wells and irregular production wells.

[0131] The production prediction model training submodule is used to train the coalbed methane production prediction model from both positive and negative directions for each type of production well, using the coalbed methane production prediction data of the sample production well as the actual coalbed methane production data of the sample production well as the label, so as to obtain a trained coalbed methane production prediction model.

[0132] As an optional implementation, the training module further includes:

[0133] The sample production well selection submodule is used to evaluate the development status of the candidate sample production wells using evaluation parameters before obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, and to obtain the evaluation results. The evaluation parameters include development time, production capacity and effective time percentage. The candidate sample production wells whose evaluation results meet the preset conditions are selected as target sample production wells.

[0134] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0135] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A data-driven method for predicting coalbed methane production, characterized in that, include: Acquire historical dynamic coalbed methane (CBM) production data and CBM production prediction data for the well to be predicted. The CBM production prediction data includes dynamic historical data and static data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes post-fracturing permeability, original gas content, geostress, and reservoir thickness. Based on the historical dynamic production data of coalbed methane, the production well type to which the production well to be predicted belongs is determined; the production well types include: conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells; conventional production wells are those whose production successively goes through three stages: a production increase period, a stable production period, and a production decrease period; single-peak production wells are those whose stable production period is less than a set threshold or whose stable production period is zero; double-peak production wells are those that include two production increase periods and two production decrease periods, and whose stable production period is less than a set threshold or whose stable production period is zero; irregular production wells are those that do not belong to the conventional production wells, single-peak production wells, or double-peak production wells; A pre-trained coalbed methane production prediction model corresponding to the type of production well is selected as the target model; the pre-trained coalbed methane production prediction model is a model trained with coalbed methane production prediction data of sample production wells as input and actual coalbed methane production data of sample production wells as labels; the pre-trained coalbed methane production prediction model adopts a bidirectional long short-term memory network model. The relevant data for predicting coalbed methane production of the well to be predicted are input into the target model to obtain the production data of the well to be predicted; the production data includes bottom hole flowing pressure, gas production, and water production. Before selecting the trained coalbed methane production prediction model corresponding to the type of production well as the target model, the method further includes training the coalbed methane production prediction model. The training process includes: acquiring coalbed methane production prediction data and actual coalbed methane production data of the sample production wells, and then preprocessing the sample data. The sample data includes coalbed methane production prediction data and actual coalbed methane production data of the sample production wells. The preprocessing of the sample data includes: Detect outliers in the sample data; Remove the outlier from the sample data; The specific method for detecting and removing outliers in sample data can be as follows: Establish a three-dimensional coordinate system using gas production Qg, water production Qw, and bottom hole flowing pressure Pw. Map the dynamic data of a single well onto the three-dimensional coordinate system based on the number of days the well has been in operation. First, calculate the distance between any two points (p, o) using the distance formula: The distance between point o and its k-th nearest point is defined as dk, and the set of all points whose distance from p to o does not exceed dk is defined as Nk(o), with the formula: Nk(o) = {p∈Z|d(p,o)≤dk} Where Z is the set of all mapping points of a single well in the three-dimensional coordinate system, and then, the absolute distance from each point p to o is determined. The formula for the absolute distance is: Adk(p,o)=max{dk,d(p,o)} The density (I) and anomaly (Abd) of this point can be expressed as: The outlier of all points is calculated using the above method and arranged in descending order. The higher the outlier, the more likely that the point is an outlier. The preprocessing of the sample data also includes filling in missing values ​​in the sample data: Identify the missing data segments in the sample data; the missing data segments include gas production and water production. Obtain the missing segment completion related data corresponding to the missing segment data; the missing segment completion related data is used to calculate the missing segment data; the missing segment completion related data includes bottom hole flowing pressure, oil pressure, tubing pressure, and dynamic fluid level; The missing segment completion data is input into the trained completion model to obtain the numerical values ​​corresponding to the missing segment data; the trained completion model is a model trained with bottom hole flowing pressure, oil pressure, tubing pressure and dynamic fluid level in normal data as input and gas production and water production in normal data as labels. The loss function for the training process is: L tol =μL s +τL d Among them, L tol Total loss; L d and L s Let μ and τ represent the dynamic loss and static loss, respectively; μ and τ represent the coefficients of the static loss and dynamic loss, respectively, and μ + τ = 1. The dynamic loss is: Here, Net represents the computation process of a bidirectional long short-term memory network using the chain rule of differentiation and the backpropagation algorithm; Net Pw This refers to the bottom hole flowing pressure predicted based on a coalbed methane production prediction model; Net Qg This represents the gas production predicted based on a coalbed methane production prediction model; Net Qw R represents the water production predicted based on the coalbed methane production prediction model; N is the total sample size; R Pw R represents the actual bottom hole flowing pressure; Qg Represents the actual gas production value; R Qw This represents the actual water production value; The static loss is: Where K is the permeability after fracturing; Gc is the original gas content; Gs is the geostress; and Rt is the reservoir thickness.

2. The method according to claim 1, characterized in that, Before selecting the trained coalbed methane production prediction model corresponding to the production well type as the target model, the method further includes training the coalbed methane production prediction model, and the training process is as follows: Obtain relevant data on coalbed methane production forecasts and actual coalbed methane production data of sample production wells; the sample production wells include conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells. For each type of production well, the coalbed methane production prediction model is trained using the coalbed methane production prediction data of the sample production well as the actual coalbed methane production data of the sample production well as the label, thus obtaining the trained coalbed methane production prediction model corresponding to the production well type.

3. The method according to claim 2, characterized in that, Before obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, the method further includes selecting sample production wells, and the selection method is as follows: The development status of the selected sample production wells is evaluated using evaluation parameters to obtain evaluation results; the evaluation parameters include development time, production capacity, and percentage of effective time. Select the candidate production wells whose evaluation results meet the preset conditions as target production wells.

4. A data-driven coalbed methane production prediction system, characterized in that, include: The data acquisition module for the production well to be predicted is used to acquire historical dynamic coalbed methane production data and coalbed methane production prediction data of the production well to be predicted. The coalbed methane production prediction data includes dynamic historical data and static data. The dynamic historical data includes historical bottomhole flowing pressure, historical gas production, historical water production, historical oil pressure, historical tubing pressure, historical dynamic fluid level, cumulative oil production, and cumulative water production. The static data includes permeability after fracturing, original gas content, geostress, and reservoir thickness. The module for determining the type of production well to be predicted is used to determine the type of production well to which the production well to be predicted belongs based on the historical dynamic production data of coalbed methane. The production well types include: conventional production wells, single-peak production wells, double-peak production wells, and irregular production wells. Conventional production wells are those whose production successively goes through three stages: a production increase period, a stable production period, and a production decrease period. Single-peak production wells are those whose stable production period is less than a set threshold or whose stable production period is zero. Double-peak production wells are those that include two production increase periods and two production decrease periods, and whose stable production period is less than a set threshold or whose stable production period is zero. Irregular production wells are those that do not belong to the conventional production wells, single-peak production wells, or double-peak production wells. The target model selection module is used to select a pre-trained coalbed methane production prediction model corresponding to the type of production well as the target model; the pre-trained coalbed methane production prediction model is a model trained with coalbed methane production prediction data of sample production wells as input and actual coalbed methane production data of sample production wells as labels; the pre-trained coalbed methane production prediction model adopts a bidirectional long short-term memory network model. The production data prediction module is used to input the coalbed methane production prediction data of the well to be predicted into the target model to obtain the production data of the well to be predicted; the production data includes bottom hole flowing pressure, gas production and water production. Before selecting the pre-trained coalbed methane production prediction model corresponding to the production well type as the target model, the system further includes training the coalbed methane production prediction model. The training process includes: acquiring coalbed methane production prediction data and actual coalbed methane production data of the sample production wells, and then preprocessing the sample data; the sample data includes coalbed methane production prediction data and actual coalbed methane production data of the sample production wells; the preprocessing of the sample data includes: Detect outliers in the sample data; Remove the outlier from the sample data; The specific method for detecting and removing outliers in sample data can be as follows: Establish a three-dimensional coordinate system using gas production Qg, water production Qw, and bottom hole flowing pressure Pw. Map the dynamic data of a single well onto the three-dimensional coordinate system based on the number of days the well has been in operation. First, calculate the distance between any two points (p, o) using the distance formula: The distance between point o and its k-th nearest point is defined as dk, and the set of all points whose distance from p to o does not exceed dk is defined as Nk(o), with the formula: Nk(o) = {p∈Z|d(p,o)≤dk} Where Z is the set of all mapping points of a single well in the three-dimensional coordinate system, and then, the absolute distance from each point p to o is determined. The formula for the absolute distance is: Adk(p,o)=max{dk,d(p,o)} The density (I) and anomaly (Abd) of this point can be expressed as: The outlier of all points is calculated using the above method and arranged in descending order. The higher the outlier, the more likely that the point is an outlier. The preprocessing of the sample data also includes filling in missing values ​​in the sample data: Identify the missing data segments in the sample data; the missing data segments include gas production and water production. Obtain the missing segment completion related data corresponding to the missing segment data; the missing segment completion related data is used to calculate the missing segment data; the missing segment completion related data includes bottom hole flowing pressure, oil pressure, tubing pressure, and dynamic fluid level; The missing segment completion data is input into the trained completion model to obtain the numerical values ​​corresponding to the missing segment data; the trained completion model is a model trained with bottom hole flowing pressure, oil pressure, tubing pressure and dynamic fluid level in normal data as input and gas production and water production in normal data as labels. The loss function for the training process is: L tol =μL s +τL d Among them, L tol Total loss; L d and L s Let μ and τ represent the dynamic loss and static loss, respectively; μ and τ represent the coefficients of the static loss and dynamic loss, respectively, and μ + τ = 1. The dynamic loss is: Here, Net represents the computation process of a bidirectional long short-term memory network using the chain rule of differentiation and the backpropagation algorithm; Net Pw This refers to the bottom hole flowing pressure predicted based on a coalbed methane production prediction model; Net Qg This represents the gas production predicted based on a coalbed methane production prediction model; Net Qw R represents the water production predicted based on the coalbed methane production prediction model; N is the total sample size; R Pw R represents the actual bottom hole flowing pressure; Qg Represents the actual gas production value; R Qw This represents the actual water production value; The static loss is: Where K is the permeability after fracturing; Gc is the original gas content; Gs is the geostress; and Rt is the reservoir thickness.

5. The system according to claim 4, characterized in that, The system further includes a training module, used to train the coalbed methane production prediction model before selecting the trained coalbed methane production prediction model corresponding to the production well type as the target model; The training module specifically includes: The sample production well data acquisition submodule is used to acquire coalbed methane production prediction data and actual coalbed methane production data of the sample production wells; the sample production wells include conventional production wells, single-peak production wells, double-peak production wells and irregular production wells. The production prediction model training submodule is used to train the coalbed methane production prediction model from both positive and negative directions for each type of production well, using the coalbed methane production prediction data of the sample production well as the actual coalbed methane production data of the sample production well as the label, so as to obtain a trained coalbed methane production prediction model.

6. The system according to claim 5, characterized in that, The training module also includes: The sample production well selection submodule is used to evaluate the development status of the candidate sample production wells using evaluation parameters before obtaining the coalbed methane production prediction data and the actual coalbed methane production data of the sample production wells, and to obtain the evaluation results. The evaluation parameters include development time, production capacity and effective time percentage. The candidate sample production wells whose evaluation results meet the preset conditions are selected as target sample production wells.

Images