A Multi-Task Learning-Based Optimization Method for Fracturing Schemes in Horizontal Coal Seams

Abstract

This invention discloses a multi-task learning method for optimizing fracturing schemes in horizontal coal seams. The method includes: collecting and statistically analyzing parameter data from the study area, and preprocessing the parameter data to obtain target data; extracting features from the target data, and randomly dividing the extracted feature data into training, validation, and test sets according to a certain proportion; creating a BP neural network model using the `newff` function, and training the BP neural network model based on the training set to obtain a target BP neural network model; and optimizing the fracturing parameters of the horizontal coal seam gas well using a particle swarm optimization algorithm based on the target BP neural network model to obtain the optimization results. This invention comprehensively considers geological, fracturing, logging, and production data. The use of multiple methods improves the generalization ability and learning performance of the neural network, helping to improve prediction accuracy and is of great significance for improving and optimizing horizontal well fracturing parameters.

Description

Technical Field

[0001] This invention belongs to the field of coalbed methane extraction technology, and in particular relates to a multi-task learning method for optimizing fracturing schemes in horizontal coal seams. Background Technology

[0002] Optimization of fracturing parameters in horizontal coalbed methane wells is a key technology for improving the efficiency and safety of coalbed methane extraction, and is of great significance for further releasing the potential of my country's coal resources and optimizing its energy structure. Currently, my country's coalbed methane industry faces challenges such as complex geological conditions and strong reservoir heterogeneity. However, fracturing designs often exhibit similarities across different geological units. Traditional fracturing construction schemes are no longer sufficient to meet the production needs of horizontal coalbed methane wells in my country. Machine learning methods have been widely applied in coalbed methane well production prediction and fracturing parameter optimization. However, the problem of optimizing fracturing parameters in horizontal coalbed methane wells involves complex variable types, high dimensionality, interrelationships between variables, and numerous constraints. Single machine learning methods may get stuck in local optima, and algorithm performance declines as the number of decision variables increases. Therefore, it is crucial to comprehensively consider reservoir geology, fracturing, and production factors to construct a multi-task artificial intelligence agent model for optimizing fracturing parameter design.

[0003] Patents with application numbers 202211572501.8 and 202111581397.4 disclose methods for optimizing fracturing parameters in horizontal wells for shale gas and shale oil, respectively. However, both only optimize fracturing parameters from the perspective of fracturing construction. Patent application number 202311407780.7 discloses a layered multi-level optimization method for fracturing parameters in horizontal wells of shale reservoir well factories, which considers geological and fracturing construction parameters. However, the geological parameters used are those interpreted from well logging, ignoring the authenticity of the well logging data itself. The differential evolution optimization algorithm used is difficult to perform multi-task learning parameter optimization. All of the above-mentioned patents have certain limitations.

[0004] In summary, current horizontal well fracturing parameter optimization has significant drawbacks. First, it often uses geological parameters interpreted from well logging for subsequent research, reducing the accuracy of the logging curves themselves and making the interpretation process susceptible to interference from uncertainties such as formula and parameter selection. Second, current coalbed methane horizontal well production fracturing parameter design does not fully integrate reservoir geological conditions, fracturing operation parameters, production data, and raw well logging data. Third, there is a lack of a multi-task AI proxy model that can integrate the above parameters and provide targeted, rapid, and accurate average gas production, maximum gas production, and long-term gas production capacity of horizontal wells. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-task learning-based method for optimizing fracturing schemes in horizontal coal seams, comprising:

[0006] Collect and statistically analyze parameter data of the study area, and preprocess the parameter data to obtain target data;

[0007] Feature extraction is performed on the target data, and the extracted feature data is randomly divided into training set, validation set and test set according to a certain ratio;

[0008] A BP neural network model is created using the newff function, and the BP neural network model is trained based on the training set to obtain the target BP neural network model.

[0009] Based on the target BP neural network model, the particle swarm optimization algorithm is used to optimize the fracturing parameters of horizontal coalbed methane wells to obtain the optimization results.

[0010] Preferably, the parameter data of the study area includes geological parameters, fracturing operation parameters, coal seam logging parameters, construction parameters, and production data;

[0011] The geological parameters include coal seam depth, coal thickness, and interbedded rock thickness.

[0012] The fracturing construction parameters include the total construction fluid volume, total sand volume, sand-carrying fluid discharge volume, proppant particle size, dosage of different particle sizes, horizontal section length, and number of fracturing stages in the horizontal well.

[0013] The logging parameters for the coal seam section include AC, DEN, GR, RD, RS, and SP;

[0014] The production data includes average gas production, peak gas production, cumulative gas production over 30 days, and cumulative gas production over 2 years.

[0015] Preferably, the process of preprocessing the parameter data includes:

[0016] The box plot method was used to identify outliers in the parameter data of the study area and remove the outlier data.

[0017] Calculate the Pearson correlation coefficient among all parameters of the horizontal well, sort the reservoir parameters according to the strength of their correlation, filter out parameters with the same correlation, and keep only one of them.

[0018] Preferably, the formula for calculating the Pearson correlation coefficient among all parameters of a horizontal well is as follows:

[0019]

[0020] Where r is the Pearson correlation coefficient, x i and y i These are sample values ​​of two variables. and is the sample mean, and n is the sample size.

[0021] Preferably, the process of feature extraction from the target data includes:

[0022] Features are extracted from geological parameters, fracturing operation parameters, and production data, including mean, median, maximum, minimum, and variance.

[0023] For well logging parameters in coal seams, MATLAB's depth convolution autoencoder-decoder is used to perform nonlinear dimensionality reduction on the well logging depth time data series and output feature parameters.

[0024] Preferably, the process of randomly dividing the extracted feature data into a training set, a validation set, and a test set according to a certain proportion includes:

[0025] Five-fold cross-validation was used to divide the dataset into five subsets. Each time, a subset was selected without repetition as the test set, and the remaining four subsets were used as the training set. This process was repeated five times.

[0026] Meanwhile, the accuracy of the model is evaluated using mean absolute error and root mean square error. The smaller the values ​​of mean absolute error and root mean square error, the smaller the difference between the true value and the predicted value, and the better the accuracy of the model's prediction results.

[0027] The formulas for the mean absolute error and the root mean square error are as follows:

[0028]

[0029]

[0030] In the formula, n represents the sample size. Indicates the predicted value. Represents the actual value.

[0031] Preferably, the process of creating a BP neural network model using the newff function includes:

[0032] Based on the newff function, a column of data is treated as a sample, with a total of m rows and n columns. The topology includes an input layer, a hidden layer, and an output.

[0033] The input layer includes geological parameters, fracturing operation parameters, and coal seam logging parameters.

[0034] The output layer contains horizontal well production data.

[0035] The hidden layer is trained using a trial-and-error method, with tansig and purelin functions used as activation functions for the hidden and output layers, respectively. The Levenberg-Marquardt method is used for training, and the initial weights and thresholds of each neuron are randomly assigned. The maximum number of training iterations is 5000, the expected training error is 0.001, and the learning rate is 0.01.

[0036] The formula expression for the newff function is as follows:

[0037] net=newff(data,label,[m,n],{'tansig','purelin'},'trainlm')

[0038] The formula for determining the number of hidden layer nodes is as follows:

[0039]

[0040] In the formula, p represents the number of input layer nodes, q represents the number of output layer nodes, and a is a constant between 1 and 10;

[0041] The formulas for the tansig and purelin functions are as follows:

[0042]

[0043] purelin(x) = x.

[0044] Preferably, before optimizing the fracturing parameters of a coalbed methane horizontal well using a particle swarm optimization algorithm, the method further includes:

[0045] Based on the target BP neural network model, the importance of geological parameters, fracturing operation parameters and coal seam logging parameters is evaluated to obtain the main factors controlling production capacity of horizontal wells, including average gas production, maximum gas production and long-term gas production capacity.

[0046] Multi-task learning is performed by training a model with minimized loss, while simultaneously optimizing the losses of multiple tasks for prediction. The minimized loss training model comprises three elements: feature extraction, task sharing, and a loss function. First, multi-task learning extracts effective features from geology, logging, fracturing, and production. Then, the joint features from multiple tasks are passed to a cross-network module, where the multi-task sharing structure and degree are automatically determined through learning. Finally, homoscedasticity uncertainty is used to measure the loss weights, partial derivatives are applied to both tasks to achieve end-to-end learning, and the loss function Lt is calculated, ultimately establishing the multi-task loss function.

[0047] Preferably, the process of optimizing the fracturing parameters of a coalbed methane horizontal well using a particle swarm optimization algorithm based on the target BP neural network model includes:

[0048] The parameter data of the study area is read as training sample data to initialize the position of each particle. With speed

[0049] Calculate the fitness value of each particle and find the individual and global optimal values;

[0050] The velocity and position of each particle are updated according to the velocity and position update formula;

[0051] Determine if the particle has reached the maximum number of iterations or if the global optimal position satisfies the minimum limit. If so, output the optimal result; otherwise, continue initializing the position of each particle. With speed

[0052] Based on the optimal output results, determine the main controlling factors of average production capacity, long-term production capacity, and maximum production capacity of horizontal wells controlled by geological, fracturing construction parameters, and logging parameters during the fracturing construction and production process of coalbed methane horizontal wells. At the same time, based on the actual fracturing construction scheme, geological conditions, and logging parameters of current coalbed methane horizontal wells, optimize the design of coalbed methane horizontal well fracturing parameters through multi-task learning of average production capacity, long-term production capacity, and maximum production capacity.

[0053] Preferably, the velocity and position of each particle are updated according to the velocity and position update formula. The formula expression for calculating the velocity and position of each particle is as follows:

[0054]

[0055]

[0056] in, This represents the position of particle i in dimension d during the nth iteration. This represents the position of particle i after the (n+1)th iteration in dimension d. The velocity of particle i in the nth iteration along dimension d. Let ω represent the velocity of particle i after the updated iteration in dimension d at the (t+1)th iteration, ω represent the inertia matrix, which indicates the degree to which the particle continues to move forward in the search space, and c1 and c2 are learning factors, representing the influence weights of the individual and the group, respectively. This represents the optimal solution for particle i in dimension d. The global optimal solution for the entire population is represented by r1 and r2, which represent the Rand function, a random number between [0,1]. The numbers represented by r1 and r2 are different.

[0057] Compared with the prior art, the present invention has the following advantages and technical effects:

[0058] This invention comprehensively integrates horizontal well logging curves, geological data, fracturing data, and coalbed methane production data. It employs a convolutional neural network to extract feature values, uses a backpropagation (BP) neural network supplemented with five-fold cross-validation, embeds a multi-task shared digital module, and uses a particle swarm optimization algorithm to optimize the fracturing parameters of coalbed methane horizontal wells. This invention emphasizes the authenticity of logging data and employs multiple types of data constraints. Compared with traditional geological methods and single artificial intelligence methods, the constructed algorithm can more easily find the global optimum during the search process, improve the generalization ability and learning performance of the neural network, and help improve prediction accuracy. It is of great significance for improving and optimizing horizontal well fracturing parameters. Attached Figure Description

[0059] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0060] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0061] Figure 2 This is a flowchart illustrating the five-fold cross-validation verification process for the training set, validation set, and test set according to an embodiment of the present invention.

[0062] Figure 3 This is a flowchart illustrating the identification of key production-controlling factors in horizontal coalbed methane wells based on a BP neural network, according to an embodiment of the present invention.

[0063] Figure 4 This is a schematic diagram of a multi-task shared digital module according to an embodiment of the present invention;

[0064] Figure 5 This is a flowchart illustrating the particle swarm optimization process for horizontal well fracturing parameters according to an embodiment of the present invention. Detailed Implementation

[0065] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0066] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0067] like Figure 1-5 As shown in the figure, this embodiment provides a multi-task learning-based optimization method for coal seam horizontal well fracturing schemes, including the following steps:

[0068] Parameter data of the study area were collected and statistically analyzed, and the parameter data were preprocessed to obtain target data. Feature extraction was performed on the target data, and the extracted feature data was randomly divided into training set, validation set, and test set according to a certain ratio. A BP neural network model was created using the newff function, and the BP neural network model was trained based on the training set to obtain the target BP neural network model. Based on the target BP neural network model, the particle swarm optimization algorithm was used to optimize the fracturing parameters of the coalbed methane horizontal well to obtain the optimization results.

[0069] More specifically, it includes the following steps:

[0070] Step 1: Compile data on the geology, fracturing operations, coal seam logging curves, and production data of the study area. Geological parameters include coal seam depth, coal thickness, and interbedded rock thickness. Fracturing operation parameters include total fluid volume, total sand volume, sand-carrying fluid discharge rate, proppant particle size, usage of different particle sizes, horizontal section length, and number of fracturing sections. Coal seam logging parameters include AC, DEN, GR, RD, RS, and SP. Production data includes average gas production, peak gas production, cumulative gas production within 30 days (maximum gas production capacity), and cumulative gas production within 2 years (long-term gas production capacity).

[0071] Step 2: During the data collection phase, variables such as geological parameters, well logging data, construction parameters, and production data were considered. Considering that there may be errors or mistakes in the actual data during the collection process, the data was preprocessed before relevant analysis work was carried out. The box plot method was used to identify outliers and remove abnormal data.

[0072] Step 3: Use formula (1) to calculate the Pearson correlation coefficient among all parameters of the horizontal well, sort the reservoir parameters according to the strength of their correlation, and filter out one of the two parameters with a strong correlation.

[0073]

[0074] Where r is the Pearson correlation coefficient, x i and y i These are sample values ​​of two variables. and is the sample mean, and n is the sample size.

[0075] Step 4: For geological parameters, fracturing operation parameters, and production data, since the data is fixed, extract features including mean, median, maximum, minimum, and variance; for well logging parameters, since it is a depth-time data sequence, use MATLAB's depth convolution auto-encoder-decoder to achieve nonlinear dimensionality reduction representation of the well logging sequence and output feature parameters.

[0076] Step 5: Randomly divide the dataset into training, validation, and test sets according to a certain ratio. The training set is used to train the machine learning model, the validation set is used to test the training degree of the model to avoid overfitting and underfitting, and the test set is used to evaluate the final effect of the model. The test set is never involved in the training and optimization process of the model. For small datasets, the model may not perform well on a randomly selected test set. In this embodiment, a five-fold cross-validation is used to divide the dataset. The dataset is divided into five subsets. Each time, a subset is selected without repetition as the test set, and the remaining four subsets are used as the training set. The training is repeated five times. In this way, all samples will have the opportunity to participate in the model training and testing, thus solving the above problem. The mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the accuracy of the model. The smaller the MAE and RMSE values, the smaller the difference between the true value and the predicted value, and the better the accuracy of the model prediction result, as shown in formulas (2)-(3), to construct the optimal partitioning model.

[0077]

[0078] In the formula, n represents the sample size. Indicates the predicted value. Represents the actual value.

[0079] Step 6: Based on the above dataset division, create a BP neural network using the newff function, as shown in formula (4). The newff function treats one column of data as a sample (m rows and n columns). Its topology includes an input layer, a hidden layer, and an output layer. The input layer contains geological, fracturing, and logging parameters, and the output layer contains horizontal well production data. The hidden layer is determined by trial and error. The number of nodes in the hidden layer is shown in formula (5). The tansig and purelin functions, as shown in formulas (6)-(7), are used as activation functions for the hidden and output layers, respectively. The Levenberg-Marquardt method is used for training. The initial weights and thresholds of each neural unit are randomly assigned. The maximum number of training iterations is 5000, the expected training error is 0.001, and the learning rate is 0.01.

[0080] net=newff(data,label,[m,n],{'tansig','purelin'},'trainlm') (4)

[0081]

[0082] In the formula, p represents the number of input layer nodes, q represents the number of output layer nodes, and a is a constant between 1 and 10.

[0083]

[0084] purelin(x) = x(7)

[0085] Step 7: Finally, the importance of geological, logging and fracturing characteristic parameters is evaluated based on the BP neural network to obtain the main factors controlling the production capacity of horizontal wells, including average gas production, maximum gas production and long-term gas production capacity.

[0086] Step 8: Multi-task learning accurately and effectively predicts by simultaneously optimizing the losses of multiple tasks. The model consists of three elements: feature extraction, task sharing, and loss function. First, multi-task learning extracts effective features from geology, logging, fracturing, and production. Then, the joint features of multiple tasks are passed to the cross-network module. Through learning, the multi-task sharing structure and degree of sharing are automatically determined, as shown in formulas (8)-(9). Finally, homoscedasticity uncertainty is used to measure the loss weight. Formulas (10)-(12) are used as auxiliary change formulas to perform partial differentiation on the two tasks, achieving end-to-end learning. The loss function Lt is calculated, and finally, the multi-task loss function is established, as shown in formula (13), to obtain the training model with the minimum loss.

[0087]

[0088] in and These are the feature maps of average capacity, maximum capacity, and long-term capacity at (i,j), respectively; α represents the cross unit, and the smaller the value, the lower the degree of sharing between tasks. When α = 0, the layer at (i,j) is a task-specific layer. The segmentation structure and degree of sharing of the network can be determined by the value of α. and These are the feature maps of average capacity, maximum capacity, and long-term capacity learned through cross-units, L. t L is the total loss function of the multi-task model. i and w i Let be the loss function and weight of the i-th task, respectively, t be the number of tasks, and σ be the population standard deviation of the normal distribution.

[0089] Step Nine: The principle of the Particle Swarm Optimization (PSO) algorithm is to simulate the changes in the position and velocity of particles in a multi-dimensional search space. Through the fusion of evolutionary information from individual learning and social learning, the particle gradually approaches the optimal position, reaching the global optimum. The entire process involves the fitness changes of each particle in the population, while also considering the influence of the global optimum. Through iteration, particles adjust their velocity and position based on their own state and the surrounding environment, ultimately achieving an optimized solution to the problem. The formulas for calculating the velocity and position of each particle are as follows...

[0090]

[0091] in, This represents the position of particle i in dimension d during the nth iteration. This represents the position of particle i after the (n+1)th iteration in dimension d. The velocity of particle i in the nth iteration along dimension d. Let ω represent the velocity of particle i after the updated iteration in dimension d at the (t+1)th iteration, ω represent the inertia matrix, which indicates the degree to which the particle continues to move forward in the search space, and c1 and c2 are learning factors, representing the influence weights of the individual and the group, respectively. This represents the optimal solution for particle i in dimension d. The global optimal solution for the entire population is represented by r1 and r2, which represent the Rand function, a random number between [0,1]. The numbers represented by r1 and r2 are different.

[0092] A detailed flowchart of the particle swarm optimization algorithm is shown below. Figure 5 As shown:

[0093] (1) Select the structure of the BP neural network, including the training set, validation set, and test set;

[0094] (2) Read the learning sample data, which includes geological parameters such as coal seam depth, coal thickness, and interbedded rock thickness; fracturing construction parameters such as total construction fluid volume, total sand volume, sand-carrying fluid discharge volume, proppant particle size, dosage of different particle sizes, horizontal section length, and number of fracturing sections; coal seam logging parameters such as AC, DEN, GR, RD, RS, and SP; and production data such as average gas production, peak gas production, cumulative gas production within 30 days (maximum gas production capacity), and cumulative gas production within 2 years (long-term gas production capacity).

[0095] (3) Initialize the position of each particle With speed

[0096] (4) Calculate the fitness value of each particle and find the individual and global optimal values;

[0097] (5) Update the velocity and position of each particle according to the velocity and position update formula;

[0098] (6) Determine whether the particle has reached the maximum number of iterations or whether the global optimal position satisfies the minimum limit;

[0099] (7) If yes, output the optimal result; otherwise, jump to (3).

[0100] Step 10: Based on the above steps, the main controlling factors for the average, long-term, and maximum production capacity of horizontal wells during coalbed methane fracturing operations can be determined, including geological parameters, fracturing parameters, and logging parameters. Simultaneously, based on the current actual fracturing schemes, geological conditions, and logging parameters of coalbed methane horizontal wells, the design of fracturing parameters for average, long-term, and maximum production capacity is optimized through multi-task learning. Taking a horizontal well as an example, the optimized fracturing scheme results in an average gas production of 2034.21 cubic meters, a peak 30-day cumulative gas production of 169,800 cubic meters, and a 2-year cumulative production of 1,343,200 cubic meters. Compared to the actual fracturing scheme, this represents a 36.5% increase in average gas production, an 18.9% increase in peak 30-day cumulative gas production, and a 22.7% increase in cumulative gas production over 2 years. See Tables 1-2 for details.

[0101] Table 1

[0102]

[0103] Table 2

[0104]

[0105]

[0106] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-task learning-based optimization method for fracturing schemes in horizontal coal seams, characterized in that, include: Collect and statistically analyze parameter data of the study area, and preprocess the parameter data to obtain target data; Feature extraction is performed on the target data, and the extracted feature data is randomly divided into training set, validation set and test set according to a certain ratio; A BP neural network model is created using the newff function, and the BP neural network model is trained based on the training set to obtain the target BP neural network model. Based on the target BP neural network model, the particle swarm optimization algorithm is used to optimize the fracturing parameters of the coalbed methane horizontal well to obtain the optimization results; Before optimizing the fracturing parameters of coalbed methane horizontal wells using the particle swarm optimization algorithm, the method further includes: Based on the target BP neural network model, the importance of geological parameters, fracturing operation parameters and coal seam logging parameters is evaluated to obtain the main factors controlling production capacity of horizontal wells, including average gas production, maximum gas production and long-term gas production capacity. Multi-task learning is performed by training a model with minimized loss, while simultaneously optimizing the losses of multiple tasks for prediction. The minimized loss training model comprises three elements: feature extraction, task sharing, and a loss function. First, multi-task learning extracts effective features from geology, logging, fracturing, and production. Then, the joint features from multiple tasks are passed to a cross-network module, where the multi-task sharing structure and degree are automatically determined through learning. Finally, homoscedasticity uncertainty is used to measure the loss weights, partial derivatives are applied to both tasks to achieve end-to-end learning, and the loss function Lt is calculated, ultimately establishing the multi-task loss function.

2. The method according to claim 1, characterized in that, The parameter data for the study area include geological parameters, fracturing operation parameters, coal seam logging parameters, construction parameters, and production data; The geological parameters include coal seam depth, coal thickness, and interbedded rock thickness. The fracturing construction parameters include the total construction fluid volume, total sand volume, sand-carrying fluid discharge volume, proppant particle size, dosage of different particle sizes, horizontal section length, and number of fracturing stages in the horizontal well. The logging parameters for the coal seam section include AC, DEN, GR, RD, RS, and SP; The production data includes average gas production, peak gas production, cumulative gas production over 30 days, and cumulative gas production over 2 years.

3. The method according to claim 1, characterized in that, The preprocessing of the parameter data includes: The box plot method was used to identify outliers in the parameter data of the study area and remove the outlier data. Calculate the Pearson correlation coefficient among all parameters of the horizontal well, sort the reservoir parameters according to the strength of their correlation, filter out parameters with the same correlation, and keep only one of them.

4. The method according to claim 3, characterized in that, The formula for calculating the Pearson correlation coefficient among all parameters of a horizontal well is as follows: Where r is the Pearson correlation coefficient, x i and y i These are sample values ​​of two variables. and is the sample mean, and n is the sample size.

5. The method according to claim 1, characterized in that, The process of feature extraction from the target data includes: Features are extracted from geological parameters, fracturing operation parameters, and production data, including mean, median, maximum, minimum, and variance. For well logging parameters in coal seams, MATLAB's depth convolution autoencoder-decoder is used to perform nonlinear dimensionality reduction on the well logging depth time data series and output feature parameters.

6. The method according to claim 1, characterized in that, The process of randomly dividing the extracted feature data into training, validation, and test sets according to a certain ratio includes: Five-fold cross-validation was used to divide the dataset into five subsets. Each time, a subset was selected without repetition as the test set, and the remaining four subsets were used as the training set. This process was repeated five times. Meanwhile, the accuracy of the model is evaluated using mean absolute error and root mean square error. The smaller the values ​​of mean absolute error and root mean square error, the smaller the difference between the true value and the predicted value, and the better the accuracy of the model's prediction results. The formulas for the mean absolute error and the root mean square error are as follows: In the formula, n represents the sample size. Indicates the predicted value. Represents the actual value.

7. The method according to claim 1, characterized in that, The process of creating a BP neural network model using the newff function includes: Based on the newff function, a column of data is treated as a sample, with a total of m rows and n columns. The topology includes an input layer, a hidden layer, and an output layer. The input layer includes geological parameters, fracturing operation parameters, and coal seam logging parameters. The output layer contains horizontal well production data. The hidden layer is trained using a trial-and-error method, with tansig and purelin functions used as activation functions for the hidden and output layers, respectively. The Levenberg-Marquardt method is used for training, and the initial weights and thresholds of each neuron are randomly assigned. The maximum number of training iterations is 5000, the expected training error is 0.001, and the learning rate is 0.

01. The formula expression for the newff function is as follows: net=newff(data,label,[m,n],{'tansig','purelin'},'trainlm') The formula for determining the number of hidden layer nodes is as follows: In the formula, p represents the number of input layer nodes, q represents the number of output layer nodes, and a is a constant between 1 and 10; The formulas for the tansig and purelin functions are as follows: 。 8. The method according to claim 1, characterized in that, Based on the target BP neural network model, the process of optimizing the fracturing parameters of coalbed methane horizontal wells using the particle swarm optimization algorithm includes: The parameter data of the study area is read as training sample data to initialize the position of each particle. With speed ; Calculate the fitness value of each particle and find the individual and global optimal values; The velocity and position of each particle are updated according to the velocity and position update formula; Determine if the particle has reached the maximum number of iterations or if the global optimal position satisfies the minimum limit. If so, output the optimal result; otherwise, continue initializing the position of each particle. With speed ; Based on the optimal output results, determine the main controlling factors of average production capacity, long-term production capacity, and maximum production capacity of horizontal wells controlled by geological, fracturing construction parameters, and logging parameters during the fracturing construction and production process of coalbed methane horizontal wells. At the same time, based on the actual fracturing construction scheme, geological conditions, and logging parameters of current coalbed methane horizontal wells, optimize the design of coalbed methane horizontal well fracturing parameters through multi-task learning of average production capacity, long-term production capacity, and maximum production capacity.

9. The method according to claim 8, characterized in that, The velocity and position of each particle are updated according to the velocity and position update formulas. The formula expressions for calculating the velocity and position of each particle are as follows: in, This represents the position of particle i in dimension d during the nth iteration. This represents the position of particle i after the (n+1)th iteration in dimension d. The velocity of particle i in the nth iteration along dimension d. Let ω represent the velocity of particle i after the updated iteration in dimension d at the (t+1)th iteration, ω represent the inertia matrix, which indicates the degree to which the particle continues to move forward in the search space, and c1 and c2 are learning factors, representing the influence weights of the individual and the group, respectively. This represents the optimal solution for particle i in dimension d. The global optimal solution for the entire population is represented by r1 and r2, which represent the Rand function, which represents a random number between [0,1]. The numbers represented by r1 and r2 are different.

Images