Coal rock gas fracturing effect prediction method and device based on machine learning

Abstract

The present application relates to a kind of coal rock gas fracturing effect prediction method and device based on machine learning, belong to coal rock gas and intelligent development technical field.The present application utilizes neural network to train and learn the key parameters influencing fracture length, and on this basis, the initial weight and threshold of neural network are globally optimized using genetic algorithm, while combining the local optimization mechanism in network training process, both improve the convergence efficiency and stability of model, and overcome the problems of neural network being easy to fall into local optimum and slow convergence speed, solve the problems of insufficient fracturing reconstruction effect prediction precision, lack of basis for feature selection, high prediction cost and the like.

Description

Technical Field

[0001] This invention relates to a method and apparatus for predicting the fracturing effect of coal and rock gas based on machine learning, belonging to the field of coal and rock gas and intelligent development technology. Background Technology

[0002] By the end of 2024, my country's coalbed methane reserves reached 70.76 trillion cubic meters, accounting for 26.2% of global reserves. Coalbed methane is gradually becoming a key link in increasing reserves and production. Hydraulic fracturing, as a crucial method for coalbed methane development, offers advantages such as "improving permeability, reducing reservoir contamination, and lowering development costs." However, during fracturing operations, interference frequently occurs due to factors such as the utilization of existing well reserves, spatial distance, and the scale of fracturing. When natural fractures are well-developed around a well or the spatial distance between wells is small, the fracture system formed by artificially created fractures in both new and old wells may connect with natural fractures, negatively impacting the production of stably generated gas wells—a phenomenon known as channeling. Specific manifestations of channeling include: significant fluctuations in bottomhole flowing pressure in old wells, severe and difficult-to-recover fluid accumulation in the wellbore, and water flooding restricting gas well productivity. In severe cases, it can even lead to sand production in old wells and casing deformation and failure. Therefore, research on channeling diagnosis and construction optimization in coalbed methane wells based on fracture network connectivity is essential and has significant practical implications.

[0003] Currently, most techniques for predicting the effects of coal and rock gas fracturing fall into two categories: regression analysis and numerical simulation. Both methods have distinct advantages and disadvantages. Regression analysis is simple and convenient to operate, but it relies on strong correlations between data and lacks the ability to handle complex situations. Numerical simulation uses finite element methods to construct multi-field coupled models, which can predict fracture formation effects under complex conditions, but its calculation process requires many geological and mechanical parameters, resulting in high computational time and cost, and it also has a certain operational threshold. Existing technologies suffer from insufficient accuracy in predicting fracturing effects, lack of basis for feature selection, and high prediction costs.

[0004] In recent years, "artificial intelligence + coal and rock gas fracturing" has gradually become an important research direction for the efficient development of coal and rock gas. Using artificial intelligence to accurately predict the effect of hydraulic fracturing is not only the key to preventing fracture flow and ensuring the safety of fracturing construction, but also the core link to optimize fracturing parameters, improve reservoir stimulation effect and formulate scientific and reasonable fracturing schemes, which is of great significance to the efficient development of coal and rock gas.

[0005] Chinese invention patent application CN120701286A discloses a method and apparatus for optimizing coalbed methane fracturing construction parameters. This method collects geological, engineering, and production data from multiple single wells in a coalbed methane development area. Based on the geological data, the multiple single wells are classified into multiple well groups. According to the parameter values ​​of the first-level engineering parameters and production data of each single well in the well group, multiple key first-level engineering parameters are determined. If at least one key first-level engineering parameter has a corresponding second-level engineering parameter, at least one key second-level engineering parameter is determined, and so on, until no corresponding sub-level engineering parameter exists for each key multi-level engineering parameter. A coalbed methane production prediction model is trained based on the key geological parameters, key multi-level engineering parameter values, and production data of each single well in the well group to predict the coalbed methane production of the single wells in the well group. This method has a complex calculation process, and errors in the preliminary steps such as feature selection and clustering can be propagated and accumulated downwards, thus affecting the accuracy of coalbed methane fracturing prediction. Due to the complexity of the calculation process, the required computational cost is also relatively high. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for predicting the fracturing effect of coal and rock gas based on machine learning, so as to solve the problems of low accuracy and high cost in the current fracturing effect prediction.

[0007] To address the aforementioned technical problems, the present invention provides, in a first aspect, a method for predicting the effects of coal and rock gas fracturing based on machine learning, the method comprising the following steps: 1) Obtain fracturing parameters related to fracture length before fracturing operations; 2) Use a neural network model as a fracturing effect prediction model. Input the fracturing parameters related to the fracture length into the fracturing effect prediction model to make predictions and obtain the corresponding fracture length, thereby judging the fracturing effect of coal and rock gas. The weights and thresholds in the fracturing effect prediction model are determined using a genetic algorithm. Supervised training is performed on the fracturing effect prediction model with determined weights and thresholds. The model with the lowest mean square error is determined, and it is judged whether the fracturing effect prediction model meets the requirements. If it does not meet the requirements, supervised training continues until the requirements are met.

[0008] In one possible implementation, the neural network model is an Elman neural network model.

[0009] In one possible implementation, the process of using a genetic algorithm to determine the weights and thresholds in the fracturing effect prediction model includes: Determine the initial population size in the genetic algorithm, randomly generate the initial population and initialize it; The fitness value of each chromosome is calculated, which is measured by the error between the prediction result of the corresponding neural network prediction model and the actual result, and the individuals in the population are ranked accordingly. Perform crossover operations according to the set crossover probability, perform mutation operations according to the set mutation probability, and update the population; The fitness of each chromosome in the updated population is calculated. When the optimal chromosome meets the predetermined requirements, the algorithm terminates. Otherwise, the population is updated again until the termination condition is met. The weights and thresholds corresponding to the optimal chromosome determined when the algorithm terminates are used as the weights and thresholds in the fracturing effect prediction model.

[0010] In one possible implementation, the fitness value is measured by the error between the predictions of the neural network model and the actual results.

[0011] In one possible implementation, determining whether the fracturing effect prediction model meets the requirements is done by comparing the prediction results of the neural network model with the actual parameters using a preset judgment criterion.

[0012] In one possible implementation, the process of obtaining fracturing parameters related to fracture length includes: Acquire coal and rock gas fracturing construction parameters and microseismic monitoring data, and perform noise reduction and normalization processing; The correlation of the data after noise reduction and normalization was quantitatively characterized, and parameters highly correlated with crack length were screened out.

[0013] In one possible approach, the quantitative characterization of correlation is achieved using Pearson correlation analysis. By calculating the Pearson correlation coefficient between each data point after noise reduction and normalization and the crack length, parameters whose absolute values ​​of the Pearson correlation coefficients are greater than a set threshold are selected as parameters that are highly correlated with the crack length.

[0014] In one possible implementation, the fracturing parameters related to the fracture length obtained include the average sand ratio, fracturing sand quantity, fracturing fluid quantity, pre-fracturing fluid quantity, extension pressure, and construction displacement.

[0015] In one possible implementation, the method further includes analyzing the performance of the neural network model with different numbers of hidden layer nodes, transfer functions, momentum factors, and learning rates according to preset evaluation criteria, and selecting the number of hidden layer nodes, transfer functions, momentum factors, and learning rates with the smallest relative errors as the number of hidden layer nodes, transfer functions, momentum factors, and learning rates of the neural network model.

[0016] To address the aforementioned technical problems, the present invention provides, in a second aspect, a machine learning-based coal and rock gas fracturing effect prediction device, comprising a processor that executes a computer program to implement the machine learning-based coal and rock gas fracturing effect prediction method of the present invention, the method comprising the following steps: 1) Obtain fracturing parameters related to fracture length before fracturing operations; 2) Use a neural network model as a fracturing effect prediction model. Input the fracturing parameters related to the fracture length into the fracturing effect prediction model to make predictions and obtain the corresponding fracture length, thereby judging the fracturing effect of coal and rock gas. The weights and thresholds in the fracturing effect prediction model are determined using a genetic algorithm. Supervised training is performed on the fracturing effect prediction model with determined weights and thresholds. The model with the lowest mean square error is determined, and it is judged whether the fracturing effect prediction model meets the requirements. If it does not meet the requirements, supervised training continues until the requirements are met.

[0017] In one possible implementation, the neural network model is an Elman neural network model.

[0018] In one possible implementation, the process of using a genetic algorithm to determine the weights and thresholds in the fracturing effect prediction model includes: Determine the initial population size in the genetic algorithm, randomly generate the initial population and initialize it; The fitness value of each chromosome is calculated, which is measured by the error between the prediction result of the corresponding neural network prediction model and the actual result, and the individuals in the population are ranked accordingly. Perform crossover operations according to the set crossover probability, perform mutation operations according to the set mutation probability, and update the population; The fitness of each chromosome in the updated population is calculated. When the optimal chromosome meets the predetermined requirements, the algorithm terminates. Otherwise, the population is updated again until the termination condition is met. The weights and thresholds corresponding to the optimal chromosome determined when the algorithm terminates are used as the weights and thresholds in the fracturing effect prediction model.

[0019] In one possible implementation, the fitness value is measured by the error between the predictions of the neural network model and the actual results.

[0020] In one possible implementation, determining whether the fracturing effect prediction model meets the requirements is done by comparing the prediction results of the neural network model with the actual parameters using a preset judgment criterion.

[0021] In one possible implementation, the process of obtaining fracturing parameters related to fracture length includes: Acquire coal and rock gas fracturing construction parameters and microseismic monitoring data, and perform noise reduction and normalization processing; The correlation of the data after noise reduction and normalization was quantitatively characterized, and parameters highly correlated with crack length were screened out.

[0022] In one possible approach, the quantitative characterization of correlation is achieved using Pearson correlation analysis. By calculating the Pearson correlation coefficient between each data point after noise reduction and normalization and the crack length, parameters whose absolute values ​​of the Pearson correlation coefficients are greater than a set threshold are selected as parameters that are highly correlated with the crack length.

[0023] In one possible implementation, the fracturing parameters related to the fracture length obtained include the average sand ratio, fracturing sand quantity, fracturing fluid quantity, pre-fracturing fluid quantity, extension pressure, and construction displacement.

[0024] In one possible implementation, the method further includes analyzing the performance of the neural network model with different numbers of hidden layer nodes, transfer functions, momentum factors, and learning rates according to preset evaluation criteria, and selecting the number of hidden layer nodes, transfer functions, momentum factors, and learning rates with the smallest relative errors as the number of hidden layer nodes, transfer functions, momentum factors, and learning rates of the neural network model.

[0025] The beneficial effects of this invention are: This invention utilizes a neural network to train and learn key parameters that affect crack length, and on this basis, uses a genetic algorithm to globally optimize the initial weights and thresholds of the neural network. At the same time, it combines the local optimization mechanism in the network training process, which not only improves the convergence efficiency and stability of the model, but also overcomes the problems of neural networks being prone to getting trapped in local optima and slow convergence speed. It solves the problems of insufficient prediction accuracy of existing fracturing transformation effects, lack of basis for feature selection, and high prediction cost. Attached Figure Description

[0026] Figure 1 This is a flowchart of the coal and rock gas fracturing effect prediction method based on machine learning according to an embodiment of the present invention; Figure 2 This is a map showing the selected target well locations according to the machine learning-based coal and rock gas fracturing effect prediction method of this invention. Figure 3 This is the input header diagram of the target well for the machine learning-based coal and rock gas fracturing effect prediction method according to an embodiment of the present invention. Figure 4 This is the output header diagram of the target well of the machine learning-based coal and rock gas fracturing effect prediction method according to an embodiment of the present invention. Figure 5This is a heatmap of the Pearson correlation coefficient between the input set of the target well and the fracture length in the machine learning-based coal rock gas fracturing effect prediction method of the present invention. Figure 6 This is a graph of the input set parameters after the target well selection for the machine learning-based coal and rock gas fracturing effect prediction method according to an embodiment of the present invention. Figure 7 This is a structural diagram of the Elman neural network model of the coal and rock gas fracturing effect prediction method based on machine learning according to an embodiment of the present invention. Figure 8 This is a schematic diagram of the genetic algorithm evolution iteration of the coal and rock gas fracturing effect prediction method based on machine learning according to an embodiment of the present invention; Figure 9 This is a graph showing the error variation of the GA-Elman neural network model in the machine learning-based coal rock gas fracturing effect prediction method of the present invention. Figure 10 This is a graph showing the prediction results of the GA-Elman neural network model for the coal and rock gas fracturing effect prediction method based on machine learning according to an embodiment of the present invention. Figure 11 This is a scatter plot of the prediction results of the unoptimized model of the machine learning-based coal and rock gas fracturing effect prediction method according to the embodiments of the present invention. Figure 12 This is a scatter plot of the prediction results of the optimized model of the machine learning-based coal and rock gas fracturing effect prediction method according to the embodiments of the present invention. Detailed Implementation

[0027] The present invention proposes a machine learning-based method for predicting the effects of coal and rock gas fracturing. This method utilizes an Elman recurrent neural network to train and learn key parameters affecting fracture length. Based on this, a genetic algorithm is used to globally optimize the initial weights and thresholds of the Elman recurrent neural network. At the same time, the local optimization mechanism in the network training process is combined to improve the convergence efficiency and stability of the model, and overcome the problems of traditional Elman recurrent neural networks being prone to getting trapped in local optima and having slow convergence speed. This solves the problems of insufficient prediction accuracy, lack of basis for feature selection, and high prediction cost of existing fracturing effects.

[0028] This invention addresses the problems of poor prediction accuracy and high prediction cost in coal and rock gas fracturing by providing a fast and efficient prediction method; while avoiding pressure channeling, it provides a safe and controllable range and maximizes the volume of fracturing stimulation.

[0029] To make the objectives, technical methods, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings. The described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] Implementation of a Machine Learning-Based Method for Predicting Coal and Rock Gas Fracturing Effects The machine learning-based method for predicting the fracturing effect of coal and rock gas in this invention first acquires fracturing parameters related to fracture length before fracturing operations. Then, a neural network model is used as the fracturing effect prediction model. The acquired fracturing parameters related to fracture length are input into this model for prediction to obtain the corresponding fracture length, thereby judging the fracturing effect of coal and rock gas. The neural network model adopts an Elman neural network model, where the weights and thresholds in the fracturing effect prediction model are determined by a genetic algorithm. The weights and thresholds optimized by the genetic algorithm are used as the initial parameters of the Elman neural network for supervised training of the fracturing effect prediction model. During training, the training samples remain unchanged, and the prediction error is reduced by iteratively updating the network parameters. When the mean square error of the model prediction result is less than a preset threshold or the maximum number of training iterations is reached, training is stopped, and the final fracturing effect prediction model is obtained. This process does not involve updating the training data; the training samples, input parameters, and corresponding labels remain unchanged. "Continuing supervised training" refers to further adjusting the network connection weights and thresholds through iterative learning under the same training data conditions, further reducing the error between the model's prediction results and the actual results.

[0031] The genetic algorithm is used only for weight and threshold optimization before training the neural network model. It obtains optimal initial network parameters through global search to improve the convergence speed and prediction accuracy of the Elman neural network. After obtaining the weights and thresholds optimized by the genetic algorithm, they are used as the initial parameters of the Elman neural network. Subsequently, supervised learning methods are used to further train the network. In this stage, the genetic algorithm is not called again; instead, the network parameters are continuously adjusted through error backpropagation.

[0032] The process of determining the weights and thresholds in the fracturing effect prediction model using a genetic algorithm includes: determining the initial population size in the genetic algorithm, randomly generating the initial population and initializing it; calculating the fitness value of each chromosome, which is measured by the error between the prediction result of the corresponding neural network prediction model and the actual result, and sorting the individuals in the population accordingly, with the fitness value measured by the error between the prediction result of the neural network model and the actual result; performing crossover operation according to the set crossover probability and mutation operation according to the set mutation probability to update the population; calculating the fitness of each chromosome in the updated population, and terminating the algorithm when the optimal chromosome meets the predetermined requirements, otherwise re-updating the population until the termination condition is met, and using the weights and thresholds corresponding to the optimal chromosome determined when the algorithm terminates as the weights and thresholds in the fracturing effect prediction model. There are two termination conditions: (1) the model prediction error reaches the preset accuracy requirement; (2) the preset maximum number of training times is reached. The model established in this invention terminates during training when its prediction error reaches the preset accuracy, such as when the mean square error reaches 0.01.

[0033] The present invention provides a method for predicting the effect of coal and rock gas fracturing based on machine learning, the flowchart of which is shown below. Figure 1 As shown, the specific steps are as follows: Step 1: Obtain the fracturing construction parameters and microseismic monitoring data of coal and gas wells in a certain area as target variables, and organize the raw data in a structured manner.

[0034] Fracturing operation parameters include: perforation depth, horizontal section length, number of perforation clusters, average sand injection intensity, average fluid injection intensity, average sand ratio, sand-fluid ratio, fracturing sand volume, fracturing fluid volume, drag-reducing water volume, operation displacement, well opening pressure, extension pressure, fracturing pressure, pump shutdown pressure, and pump shutdown pressure drop.

[0035] Data analysis was performed on the fracturing construction parameters, and the value range of different parameters was determined. At the same time, the above fracturing construction parameters were divided into three categories: fracturing design parameters, construction parameters, and pressure response parameters.

[0036] Step 2: Data Preprocessing. To prevent human-induced anomalies from affecting model training, normalization is performed to eliminate the influence of units of measurement. Specifically, this invention draws box plots and identifies outliers, then performs noise reduction and normalization on both the input and output sets to improve data quality and thus increase model prediction accuracy.

[0037] Step (2.1): Using the box plot method, a box plot is drawn based on the fracturing construction parameters and microseismic monitoring data obtained in Step 1. Outliers in different parameters are identified and their causes determined. The causes of the outliers are then denoised. If the outlier is due to an input error, the data point is deleted. Errors may occur during the on-site data collection or recording process, causing input parameters to deviate from reality. For example, errors in recording the displacement, sand addition, or fracturing fluid usage during fracturing construction can lead to inconsistencies between the model input data and the actual construction conditions, resulting in prediction bias. Here, identifying outliers confirms whether they are accurate data. If it is due to human error in statistical analysis, the outlier is deleted or revised.

[0038] Step (2.2): Normalize the denoised data to eliminate the dimensional influence between indicators.

[0039] Step 3: Dimensionality reduction is performed on the input set, and Pearson correlation analysis is used to screen key factors.

[0040] Step (3.1): Using Pearson correlation analysis, the data after noise reduction and normalization in Step 2 is quantitatively characterized for correlation. Parameters highly correlated with the target variable are selected and prioritized based on the Pearson correlation coefficient. This coefficient is a statistical indicator that measures the degree of linear correlation between two variables, with a value ranging from -1 to 1: a coefficient close to 1 indicates a high positive correlation; close to -1 indicates a high negative correlation; and close to 0 indicates almost no linear relationship.

[0041] Step (3.2): Using the correlation coefficient chart calculated in step (3.1), parameters with a Pearson correlation coefficient greater than 0.5 or less than -0.5 are selected as the input set of the final model to achieve the purpose of data dimensionality reduction. The selected input set is more significantly correlated with the results of microseismic monitoring, which can significantly improve the prediction speed and accuracy of the prediction model.

[0042] Step 4: Establish a GA-Elman neural network prediction model.

[0043] Step (4.1): Optimize the Elman neural network model using a genetic algorithm, determine the initial population size in the genetic algorithm, randomly generate the initial population and initialize it, and set optimization parameters such as crossover probability, mutation probability and termination generation.

[0044] Step (4.2): Calculate the fitness value for each chromosome, which is measured by the error between the prediction result and the actual result of the corresponding Elman neural network prediction model, and rank the individuals in the population accordingly. Specifically, the fitness value is characterized by the sum of squares of the errors between the actual output and the expected output.

[0045] Step (4.3): Perform the crossover operation according to the crossover probability set in step (4.1). Chromosomes that do not undergo crossover will self-replicate. To maintain population diversity, perform the mutation operation according to the set mutation probability, thereby ensuring the effectiveness of the algorithm.

[0046] Step (4.4): Calculate the fitness value of each chromosome and add the newly generated chromosome to the original population to form a new population. Compare the fitness of each chromosome. When the optimal chromosome meets the predetermined requirements, the algorithm terminates; otherwise, repeat steps (4.2) to (4.4) until the algorithm termination condition is met.

[0047] Step (4.5): Considering that the network model includes other parameters besides weights and thresholds (such as the number of hidden layer nodes, transfer function, momentum factor, and learning rate), this invention analyzes the performance of the neural network model with different numbers of hidden layer nodes, transfer functions, momentum factors, and learning rates according to preset evaluation criteria. The hidden layer node count, transfer function, momentum factor, and learning rate with the smallest relative error are selected as the hidden layer node count, transfer function, momentum factor, and learning rate of the neural network model. The number of hidden layer nodes in the model is determined, and Logsig is selected as the transfer function of the neural network model. Parameters such as the momentum factor, learning rate, and maximum number of iterations in the model are set.

[0048] In this invention, the genetic algorithm is mainly used to optimize the connection weights and thresholds of the Elman neural network model to obtain better initial network parameters, thereby improving the model's convergence speed and prediction accuracy. Parameters such as the number of hidden layer nodes, transfer function, learning rate, and momentum factor are neural network structural parameters and training control parameters, and are not determined by the genetic algorithm. Specifically, the number of hidden layer nodes is determined by setting different numbers of nodes to establish the prediction model and based on the changes in the model's prediction error; the transfer function is selected based on the nonlinear mapping characteristics of the Elman neural network; the learning rate and momentum factor are determined through testing multiple parameter combinations, using the mean square error and prediction accuracy during model training as evaluation indicators to determine the optimal parameter combination. Once these parameters are determined, they remain unchanged, and then the genetic algorithm is used to optimize the initial weights and thresholds of the network, ultimately forming the GA-Elman fracturing effect prediction model.

[0049] In this invention, the evaluation of model prediction performance is not based on a single evaluation index, but rather on a comprehensive evaluation system consisting of the coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error. The optimal prediction model is selected based on the largest coefficient of determination and the smallest root mean square error, mean absolute error, and mean absolute percentage error. Using these performance metrics as evaluation standards, the model performance is analyzed for different numbers of hidden layer nodes, and the number of hidden layer nodes with the smallest relative error is selected as the model input value. The same evaluation criteria are used to select parameters such as the transfer function, momentum factor, and learning rate in the neural network model.

[0050] Step 5: Conduct supervised training on the established neural network model to ensure its robustness and accuracy.

[0051] The purpose of step 4 is to obtain better initial network parameters, and the purpose of step 5 is to further train and optimize and determine the final prediction model. Step (5.1): Use the parameters selected in step (3.2) as the input set to perform supervised training on the initial weights and thresholds optimized by the genetic algorithm in step 4, and reduce the prediction error by iteratively updating the network parameters.

[0052] The purpose and optimization mechanism of step (5.1) are different from those of steps (4.1) to (4.4).

[0053] Steps (4.1) to (4.4) constitute the genetic algorithm optimization process. Its main function is to utilize the global search capability of the genetic algorithm to find optimal initial connection weights and thresholds for the Elman neural network within a large parameter space. The optimization target is the initial network parameters, aiming to improve the convergence speed and prediction accuracy of subsequent neural network training. Step (5.1) is the supervised training process for the Elman neural network, where the optimization target is the network parameters initialized by the genetic algorithm. This process employs the backpropagation method, further adjusting the network connection weights and thresholds based on the error between the predicted and actual results, gradually reducing the model's prediction error and achieving a stable state. These two steps correspond to the global optimization and local training stages, respectively. The former is used to avoid the network getting trapped in local optima during training, while the latter is used to achieve fine-tuning of the model parameters.

[0054] Step (5.2): Output the model result with the lowest error value from step (5.1), compare it with the model prediction results, and compare the scatter plot of the prediction results with the actual parameters to obtain the coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error of the prediction results. Evaluate the model prediction effect through these four output indicators and obtain the best prediction model. The accuracy of the model prediction results is comprehensively evaluated through the above four indicators to prevent the influence of the bias of a single indicator.

[0055] Step 6: During the fracturing operation, optimize the fracturing process and construction parameters for the newly implemented well.

[0056] By utilizing a trained GA-Elman neural network model, fracturing parameters are input into the model before fracturing operations to predict the fracturing effect and thus the fracturing scale of each section of the horizontal well. This prevents excessive fracturing scale from causing cross-flow between adjacent wells, while simultaneously increasing reservoir stimulation volume and optimizing fracturing operation parameters to achieve cost reduction and efficiency improvement while ensuring production safety. The predictive model established in this invention aims to predict the fracturing scale, which can be calculated using specific parameters such as fracture length, fracture width, and fracture height as the fracturing scale.

[0057] This invention is applicable to various hydraulically fracturing coal and gas horizontal wells, guiding the optimization design of fracturing processes and construction parameters to maximize reservoir modification and increase production while avoiding hydraulic channeling.

[0058] The following section uses the coal-rock gas reservoir in region A as an example to explain in detail the implementation process of the coal-rock gas fracturing effect prediction method based on machine learning in this invention.

[0059] Step 1: Based on well logging, seismic, and fracturing data, obtain coal and rock gas fracturing construction parameters and microseismic monitoring data as target variables, and then organize the raw data in a structured manner.

[0060] The gas field in region A has good overall coal seam porosity and permeability, with a porosity of 3-5%, permeability of 0.02-0.08 mD, reservoir pressure of 9-11 MPa, and reservoir temperature of 38 ℃, classifying it as a low-temperature, low-pressure coal-gas reservoir. With the continuous upgrading and iteration of effective propped fracturing technology, breakthroughs have been achieved in field tests of effective propped fracturing technology in this region. This technology has been extended to high-stress retention zones, and simultaneous experiments have been conducted to expand the fracturing scale, increase the proportion of pre-fracturing fluid, and increase the size of proppant particles in high-stress zones. Since the geological parameters, coal structure, and mineralization are not significantly different within the high-stress retention zones, the influence of geological parameters on the fracturing results is ignored. The target well location map is shown below. Figure 2 As shown, fracturing construction parameters and microseismic monitoring data were obtained, and the raw real data were structured and processed as follows: The fracturing operation parameters include the following: perforation depth, horizontal section length, number of perforation clusters, average sand injection intensity, average fluid injection intensity, average sand ratio, sand-fluid ratio, fracturing sand volume, fracturing fluid volume, drag-reducing water volume, operation displacement, well opening pressure, extension pressure, fracturing pressure, pump shutdown pressure, and pump shutdown pressure drop.

[0061] Data analysis was conducted on the fracturing construction parameters to determine the value range of different parameters. At the same time, the 16 fracturing construction parameters were divided into three categories: fracturing design parameters, construction parameters, and pressure response parameters.

[0062] Step 2: Data preprocessing. Noise reduction and normalization are performed on the input and output sets respectively to improve data quality and thus increase the model's prediction accuracy.

[0063] To prevent outliers from affecting the results, a box plot method was used. This involved analyzing the range of fracturing parameters and microseismic monitoring data obtained in step 1 and plotting a box plot to identify outliers. Outliers were then filtered, and noise reduction was applied to the truly outliers. The resulting box plot is shown below. Figure 3 , Figure 4 As shown. (Through) Figure 3 It was found that, apart from the rupture pressure, construction flow rate, and pump shutdown pressure, there were few outliers among the other parameters, while the values ​​of injection intensity and extension pressure ranged widely, and the values ​​of parameters such as average sand ratio and sand-liquid ratio were low. Figure 4 The average fracture length of Y21-TP4 and Y21-TP12 is only 553m. In contrast, the fracture length of Y21-TP3 is more ideal, with most fracture lengths concentrated within 680~750m. However, its fracture length span is large, and there are no outliers in the fracture length of the three wells.

[0064] Because different parameters in fracturing operations have different units of measurement and vary significantly in magnitude (e.g., the order of magnitude difference between displacement and fluid volume is hundreds of times), the network model must adapt to different distributions in each iteration, greatly reducing the training speed and effectiveness. Therefore, to eliminate the influence of different dimensions between indicators, a normalization method is used to preprocess the sample data.

[0065] Step 3: Dimensionality reduction of the input set, establishment of a correlation chart of fracturing parameters, and screening of key factors using Pearson correlation analysis.

[0066] Step (3.1) uses Pearson correlation analysis to quantitatively characterize the correlation of the data volume after noise reduction and normalization in step 2, selects parameters that are highly correlated with the target variable, and prioritizes them according to the Pearson correlation coefficient, as follows: To comprehensively and accurately reveal the relationship between fracturing parameters and fracturing effects, Pearson correlation analysis from statistical analysis was used to quantitatively characterize the linear correlation between multiple variables. This method can efficiently screen parameters highly correlated with the target variable while preserving the original characteristics of the data, and prioritize them according to the Pearson correlation coefficient. This coefficient is a statistical indicator that measures the degree of linear correlation between two variables, with a value ranging from -1 to 1: when the coefficient is close to 1, it indicates a high positive correlation; close to -1, it indicates a high negative correlation; and close to 0, it indicates that there is almost no linear relationship between them.

[0067] Statistical results are as follows Figure 5 As shown, it can be found that parameters such as fracturing sand volume, fracturing fluid volume, and discharge volume all have a positive correlation with crack length. Among them, the construction discharge volume has the strongest correlation with crack length, reaching 0.75, which proves that the change in construction discharge volume has the greatest impact on crack propagation. On the other hand, the fracturing pressure and the pressure drop after pump shutdown have a negative correlation with crack length, that is, the higher the fracturing pressure, the shorter the crack.

[0068] Step (3.2) uses the correlation coefficient chart calculated in step (3.1) to select parameters with a Pearson correlation coefficient greater than 0.5 as the final input set, achieving the purpose of data dimensionality reduction. The selected input set has a more obvious correlation with the results of microseismic monitoring, which can significantly improve the prediction speed and accuracy of the prediction model, as detailed below: Subsequently, based on the correlation coefficient chart calculated above, six parameters with a Pearson correlation coefficient greater than 0.5 or less than -0.5 with fracture length were selected as input layers for predicting the effectiveness of coal and rock gas fracturing. These parameters are: average sand ratio, fracturing sand quantity, fracturing fluid quantity, pre-fracturing fluid quantity, extension pressure, and construction displacement. The correlation coefficient chart of the input set is shown below. Figure 6 As shown.

[0069] Step 4: Establish a GA-Elman neural network prediction model.

[0070] Traditional Elman neural networks are characterized by their simple structure and low computational cost. Because they have built-in layers, these models are sensitive to historical data and highly adaptable to small training sets. The network model structure diagram is shown below. Figure 7 As shown. However, the traditional Elman neural network model is prone to getting trapped in local optima, so a genetic algorithm is chosen to optimize the core network design. Specifically, as shown in steps (4.1) to (4.4): Step (4.1) uses a genetic algorithm to optimize the Elman neural network model, determines the initial population size in the genetic algorithm, randomly generates the initial population and initializes it, and sets optimization parameters such as crossover probability, mutation probability and termination generation.

[0071] Step (4.2) calculates the fitness value for each chromosome, which is measured by the error between the corresponding predicted and actual results, and ranks the individuals in the population accordingly. Specifically, it calculates the sum of squares of the errors between the actual and expected outputs: Step (4.3) performs the crossover operation according to the crossover probability set in step (4.1). Chromosomes that do not undergo crossover will self-replicate. To maintain population diversity, mutation operation is performed according to the set mutation probability, thereby ensuring the effectiveness of the algorithm.

[0072] Step (4.4) calculates the fitness value of each chromosome and adds the newly generated chromosome to the original population to form a new population. Compare the fitness of each chromosome. When the optimal chromosome meets the predetermined requirements, the algorithm terminates; otherwise, repeat steps (4.2) to (4.4) until the algorithm termination condition is met.

[0073] Step (4.5) determines the number of hidden layer nodes in the model, selects Logsig as the transfer function of the neural network model, and sets parameters such as momentum factor, learning rate, and maximum number of iterations. Specifically: First, the number of hidden layer nodes in the model was designed. Using the coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error as evaluation criteria, the model performance with different numbers of hidden layer nodes was analyzed. The number of hidden layer nodes with the smallest relative error was selected as the model input value. The transfer function, momentum factor, and learning rate in the neural network model were then selected using the same evaluation criteria. By adjusting the parameters, the number of hidden layer nodes was determined to be 6, the model transfer function was chosen as Logsig, the momentum factor was set to 0.9, the learning rate to 0.02, the training accuracy to 0.001, and the maximum number of iterations to 1000. At this point, the model achieved optimal overall performance.

[0074] Step 5: Conduct supervised training on the established model to ensure its robustness and generalization ability, thereby improving its applicability.

[0075] Step (5.1) uses the parameters selected in step (3.2) as the input set to perform supervised training on the prediction model set in step 4, finds the optimal fitness value of the model, and inputs the weights and thresholds at the optimal fitness value into the neural network model for further training. As the number of iterations increases, the model with the lowest mean square error is found, as follows: The model with predefined parameters is trained, and the training process is as follows: Figure 8 As shown, it can be observed that the fitness value decreases with the increase of the number of generations. Before the 20th generation, the fitness decreases rapidly, indicating that the model has a good early evolution effect. As the number of generations continues to increase, the fitness decreases more gradually and reaches its minimum value in the 33rd generation, at which time the fitness is 0.04. Subsequently, as the number of generations increases, the fitness increases slightly but the overall change is not significant. Therefore, it is believed that the model reaches its optimal solution when it evolves to the 33rd generation, at which point the model's prediction accuracy is the highest.

[0076] The optimized weights and thresholds are then fed into the neural network model for further training. The training process is as follows: Figure 9As shown in the figure, the error variation graph of the GA-Elman neural network model reveals that, under the optimization of the genetic algorithm, the Elman neural network model has a large rate of decrease in mean square error in the first 6 iterations, indicating that its accuracy is gradually increasing. However, as the number of iterations increases, the rate of decrease in mean square error slows down, and the model error is reduced to the lowest level in the 9th iteration.

[0077] Step (5.2) outputs the model result with the lowest error value from step (5.1), compares it with the model prediction results, and compares the scatter plot of the prediction results with the actual parameters to obtain the determination coefficient, root mean square error, mean absolute error, and mean absolute percentage error of the prediction results. The model prediction performance is evaluated through these four output indicators. If the model prediction performance is poor, supervised training is performed again. If the model prediction performance is good, the best-performing prediction model is obtained, as detailed below: The training results are analyzed, and the training results are as follows: Figure 10 , 11 As shown in Figure 12. (Through) Figure 10 By comparing the deviation distribution between the predicted and actual values ​​of the two prediction models, it was found that in the first half of the training set, the two models showed a large deviation, but the data trends were generally consistent. In the latter half of the training set and the validation set, the Elman neural network, due to its reliance on backpropagation to adjust internal weights and thresholds, was prone to getting trapped in local optima, resulting in poor generalization ability and a large error compared to the actual crack length. In contrast, the model optimized by the genetic algorithm found the optimal weights and thresholds through global adaptive search, significantly improving the model's robustness and generalization ability. The mean absolute percentage error decreased from 7.63% to 2.19%, especially in the validation set, where the error between the predicted and actual values ​​was only 1.1%.

[0078] Figure 11 , 12 To compare the scatter plots of the prediction results from the two models, the more concentrated the scatter points are on a straight line with a slope of 1, the closer the predicted values ​​are to the actual values. Figure 11 , 12 It can be observed that the unoptimized Elman neural network model has strong dispersion in its prediction results, with the data being far from the regression line and the coefficient of determination between it and the actual crack length being only 0.74. However, after optimization by the genetic algorithm, the error between the predicted value and the actual value is reduced, and the fitting curve is closer to the y=x curve, indicating that the optimized model has high stability and its fitting accuracy is also improved. The coefficient of determination between the predicted result and the actual parameter is increased to 0.97.

[0079] Step 6: Before fracturing, optimize the fracturing operation parameters of a well to be evaluated.

[0080] By using a trained GA-Elman neural network model, fracturing parameters are input into the model before fracturing operations to predict the fracturing scale of each section of the horizontal well. This prevents excessive fracturing scale from causing cross-flow between adjacent wells. Simultaneously, it can increase reservoir stimulation volume and optimize fracturing operation parameters while ensuring production safety, thereby achieving the goal of cost reduction and efficiency improvement. Details are as follows: Before deploying new well locations and carrying out fracturing operations, it is necessary to pay attention to the problem of reduced production rate of adjacent wells due to the presence of high-yield old wells in the vicinity. By inputting the six key parameters in the design, such as extended pressure, average sand ratio, single-stage sand volume, single-stage fluid volume, pre-flush fluid volume, and construction discharge rate, the fracturing effect can be predicted, and the fracturing scale (such as hydraulic fracture length, fracture width, or fracture height) can be predicted. This will guide the deployment well spacing of new wells and optimize fracturing operation parameters.

[0081] Therefore, the evaluation parameters of this invention are comprehensive and the method is simple and fast. It can accurately predict the fracturing effect between coal and rock gas wells and provide technical guidance for optimizing fracturing construction to maximize gas well production capacity.

[0082] Implementation of a Machine Learning-Based Coal and Rock Gas Fracturing Effect Prediction Device This embodiment of the machine learning-based coal and rock gas fracturing effect prediction device includes a processor. The processor executes a program in the memory to implement the machine learning-based fracturing effect prediction method described above. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. The memory can be internal memory of the random access memory (RAM) type. The processor and memory can be integrated into one or more independent circuits or hardware, such as an application-specific integrated circuit (ASIC). It should be noted that the computer program in the aforementioned memory can be implemented as a software functional unit and sold or used as an independent product, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the implementation method of the coal and rock gas fracturing effect prediction method based on machine learning of the present invention.

[0083] The machine learning-based prediction method for coal and rock gas fracturing effects of this invention first clarifies that fracturing construction parameters can be divided into three categories: fracturing design parameters, construction parameters, and pressure response parameters. Then, based on the Pearson correlation coefficient method, six parameters with the highest correlation to fracture length are selected: construction flow rate, extension pressure, pre-flush fluid volume, fracturing fluid volume, fracturing sand volume, and average sand ratio. Finally, the determination coefficient of the optimized neural network model using a genetic algorithm is increased from 0.74 to 0.97 compared to the unoptimized model, and the mean absolute percentage error is reduced from 7.63% to 2.19%. This indicates that the optimized GA-Elman neural network model has significantly improved robustness and generalization ability, and a marked improvement in the prediction accuracy for fracturing scale.

[0084] This invention predicts the fracturing scale of new wells in old areas, saving 80,000-100,000 yuan per well in microseismic monitoring costs. The prediction results are then used to optimize the fracturing scale, reducing the average cost per fracturing stage by 60,000-100,000 yuan, a saving of 9.7%. Simultaneously, the model prediction results are fully considered during well placement, improving the production rate of adjacent old wells and helping 15 surrounding old wells maintain their current daily gas production levels, indirectly increasing daily gas production by 20,000-35,000 cubic meters per day, with an estimated increase in revenue of 960,000-1,680,000 yuan.

[0085] This invention can be widely applied to the development of coal and gas reservoirs in different basins, strata, and types. It has broad application prospects for tapping the remaining gas in old coal and gas fields, improving the utilization of single wells, and implementing development plans. It will strongly support my country's natural gas supply and green and low-carbon energy transformation.

Claims

1. A method for predicting the effect of coal and rock gas fracturing based on machine learning, characterized in that, The method includes the following steps: 1) Obtain fracturing parameters related to fracture length before fracturing operations; 2) Use a neural network model as a fracturing effect prediction model. Input the fracturing parameters related to the fracture length into the fracturing effect prediction model to make predictions and obtain the corresponding fracture length, thereby judging the fracturing effect of coal and rock gas. The weights and thresholds in the fracturing effect prediction model are determined using a genetic algorithm. Supervised training is performed on the fracturing effect prediction model with determined weights and thresholds. The model with the lowest mean square error is determined, and it is judged whether the fracturing effect prediction model meets the requirements. If it does not meet the requirements, supervised training continues until the requirements are met.

2. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 1, characterized in that, The neural network model described uses the Elman neural network model.

3. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 1, characterized in that, The process of determining the weights and thresholds in the fracturing effect prediction model using a genetic algorithm includes: Determine the initial population size in the genetic algorithm, randomly generate the initial population and initialize it; The fitness value of each chromosome is calculated, which is measured by the error between the prediction result of the corresponding neural network prediction model and the actual result, and the individuals in the population are ranked accordingly. Perform crossover operations according to the set crossover probability, perform mutation operations according to the set mutation probability, and update the population; The fitness of each chromosome in the updated population is calculated. When the optimal chromosome meets the predetermined requirements, the algorithm terminates. Otherwise, the population is updated again until the termination condition is met. The weights and thresholds corresponding to the optimal chromosome determined when the algorithm terminates are used as the weights and thresholds in the fracturing effect prediction model.

4. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 3, characterized in that, The fitness value is measured by the error between the predictions of the neural network model and the actual results.

5. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 1, characterized in that, To determine whether the fracturing effect prediction model meets the requirements, a preset judgment criterion is used to compare the prediction results of the neural network model with the actual parameters.

6. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 1, characterized in that, The process of obtaining fracturing parameters related to fracture length includes: Acquire coal and rock gas fracturing construction parameters and microseismic monitoring data, and perform noise reduction and normalization processing; The correlation of the data after noise reduction and normalization was quantitatively characterized, and parameters highly correlated with crack length were screened out.

7. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 6, characterized in that, The correlation was quantitatively characterized using Pearson correlation analysis. The Pearson correlation coefficient between each data point after noise reduction and normalization and the crack length was calculated. Parameters whose absolute values ​​of the Pearson correlation coefficients were greater than a set threshold were selected as parameters that were highly correlated with the crack length.

8. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 6, characterized in that, The fracturing parameters related to the fracture length obtained include average sand ratio, fracturing sand volume, fracturing fluid volume, pre-fracturing fluid volume, extension pressure, and construction discharge rate.

9. The method for predicting the effect of coal and rock gas fracturing based on machine learning according to claim 2, characterized in that, The method also includes analyzing the performance of the neural network model under different hidden layer node numbers, transfer functions, momentum factors and learning rates according to preset evaluation criteria, and selecting the hidden layer node numbers, transfer functions, momentum factors and learning rates with the smallest relative errors as the hidden layer node numbers, transfer functions, momentum factors and learning rates of the neural network model.

10. A machine learning-based device for predicting the effect of coal and rock gas fracturing, characterized in that, The system includes a processor that executes a computer program to implement the steps of the machine learning-based method for predicting the effects of coal and rock gas fracturing as described in any one of claims 1 to 9.

Images