Drilling extraction parameter double-layer optimization method based on numerical simulation and deep learning

Abstract

The invention provides a drill hole extraction parameter double-layer optimization method based on numerical simulation and deep learning, relates to the technical field of gas extraction parameter optimization, solves the technical problem of insufficient adjustment precision of coal seam gas bedding extraction process parameters, and comprises the following steps: establishing a numerical model considering multiple effects according to specific problems; inputting fracturing parameters to simulate and generate a gas concentration prediction data set; performing hard constraint mapping on the prediction data set to form a simulation data set, inputting the simulation data set into a deep learning model in combination with an empirical data set so as to reversely optimize the numerical model, and correcting the fracture form; on the basis of the numerical model, an EDFM model is adopted, extraction process parameters are input for numerical simulation to obtain a gas extraction quantity prediction set, a new simulation sample set is obtained through hard constraint mapping, an experience sample set is combined to be input into a deep learning model for training, finally, optimal extraction process parameters are obtained, and the prediction precision of the whole model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas extraction parameter optimization, and specifically relates to a double-layer optimization method for drilling extraction parameters based on numerical simulation and deep learning. Background Art

[0002] Gas extraction technology is an effective means of preventing and controlling gas disasters in mines. Due to the anisotropy of the physical and mechanical properties of coal reservoirs and the complexity of the on-site environment for extraction, many process parameters involved in gas extraction technology should be analyzed based on specific circumstances. Artificial determination based solely on practical experience will result in potential risks, low efficiency, imbalance in engineering input-output, and difficulty in controlling the extraction effect, making it difficult to achieve the maximum extraction technology efficiency.

[0003] At present, many gas extraction parameter control technologies have involved many methods such as machine learning, deep learning, and numerical simulation, but there are still the following shortcomings: Since machine learning, deep learning and other algorithms are based on data to learn the potential mapping relationship in nonlinear systems, sparse data sets often make it difficult for the model to extract deeper data features and laws. In addition, many studies have not fully integrated physical mechanisms such as coal seam storage properties, gas diffusion-seepage, and mutual coupling of multiple physical fields in gas extraction, resulting in machine learning and other models can only learn surface data relationships. The reliability, interpretability and robustness of the results are poor, and the learning accuracy needs to be further improved. Summary of the invention

[0004] In order to overcome the difficulties of artificial setting of existing gas extraction technical parameters and difficulty in determining the extraction effect, and to further improve the gas extraction efficiency of the coal seam in the tunnel, the present invention aims at the technical problem of insufficient adjustment accuracy of coal seam gas extraction process parameters, and provides a two-layer optimization method for drilling extraction parameters based on numerical simulation and deep learning.

[0005] The design concept of the present invention is as follows: from the two aspects of numerical simulation / calculation and deep learning algorithm, the present invention adopts the historical gas concentration distribution to correct the geological model of the coal seam, uses the historical gas extraction volume to optimize and adjust the gas extraction process parameters of the coal seam, and applies the numerical simulation method to construct a physical model that fits the reality. Combined with the advantage of deep learning in extracting high-level abstract features of data, it learns the effective information in the physical model and the actual distribution laws in empirical cases, thereby effectively integrating the physical information in gas extraction into the prediction model, so that the obtained results are consistent with the ideal laws and fit the actual distribution. The good generalization ability of the deep learning algorithm can also improve the prediction reliability of the gas extraction effect.

[0006] The present invention is achieved through the following technical solutions:

[0007] The two-layer optimization method of drilling and extraction parameters based on numerical simulation and deep learning includes the following steps:

[0008] S1. Establish initial numerical model A1:

[0009] Collect geological parameters and engineering parameter data, determine the fracturing parameters to be inverted, the gas extraction process parameters to be inverted and their empirical ranges, and generate the input data set U through LHS sampling i and U k ; Set geometric model parameters, boundary conditions, initial conditions, and assumptions, determine the mathematical model of gas drilling extraction based on gas desorption and adsorption characteristics, molecular diffusion, seepage characteristics, and slippage effects, import various parameters, conditions, and geometric models into COMSOL or FLUENT numerical simulation software, import the fracture propagation model, and establish the initial numerical simulation model A1;

[0010] S2, the simulation sample set I1 and the experience sample set I2 that form the deep learning model:

[0011] After numerical simulation, the gas concentration distribution data set c is generated i * , for the dataset c i * Perform hard constraint mapping of domain knowledge and generate a gas concentration dataset c that complies with physical law constraints in the local constraint block. i , input data set U i and gas concentration dataset c i Combined into simulated sample set I1: (U i , c i ); Based on past historical experience, we form an experience sample set I2: (E j , e j ), E j is the range of geological parameters and fracturing parameters in historical data, e j is the gas concentration distribution in historical data;

[0012] S3. Build a deep learning model B1 based on the simulated sample set I1 and the empirical sample set I2, and improve the loss function:

[0013] A fully connected neural network is used, and the initial number of hidden layers, number of units, optimization algorithm, learning rate and activation function are set. The hyperparameters are continuously optimized iteratively during training, and the loss function L is:

[0014] L=λL M +μL E ; (1)

[0015] In formula (1), L M is the simulated sample loss, λ is the simulated sample loss weight coefficient; L Eis the experience sample loss, μ is the experience sample loss weight coefficient; λ and μ are both constants, which are experimentally adjusted during the training of the deep learning model B1;

[0016] Simulated sample loss L M According to the partial differential equation in the hard constraint mapping, it is obtained as shown in formula (2):

[0017]

[0018] In formula (2), is the deviation of the ith partial differential equation, are all the variables in the ith partial differential equation system, N f is the number of partial differential equations;

[0019] Experience sample loss L E Calculate according to formula (3):

[0020]

[0021] In formula (3), e j , are the observed value and predicted value of the empirical sample set, respectively, and n is the number of empirical samples;

[0022] S4, model training and verification, to obtain the optimal fracturing parameter combination: the simulation sample set I1 and the experience sample set I2 are combined into a training set and a verification set, the training set is input into the deep learning model B1 for training to invert the input parameters, and the Markov chain Monte Carlo sampling method (MCMC method) is used to screen the optimal candidate set in each iteration. After the training is completed, the verification set is used for verification, and finally the optimal fracturing parameter combination is inverted;

[0023] S5. Correct the fracture morphology, integrate the physical laws in gas extraction, and prepare for numerical simulation: input the optimal fracturing parameter combination obtained in step S4 into the initial numerical model A1 established in step S1, generate the numerical model A2 after correcting the fracture morphology, and use the EDFM model (embedded discrete fracture model) to connect the numerical model A2 to form a fracture network that is closer to physical reality. The numerical model at this time is more in line with reality;

[0024] S6. Construct simulated sample set I3: Input data set U k Input the numerical simulation software and use the numerical model A2 established in step S5 to simulate and generate the pure amount of gas extraction. The same method as step S2 is used to extract the pure amount of gas. Perform hard constraint mapping of domain knowledge to generate a gas extraction scalar distribution q that meets physical constraints k, combined input dataset U k And the simulation results of gas extraction scalar distribution q k For the simulated data set I3: (U k ,q k );

[0025] S7, build deep learning model B2, model training and verification, obtain the optimal gas extraction process parameters after inversion: Use the same method as step S3 to build deep learning model B2 and simulate data set I3: (U k ,q k )Combined with empirical data set I4: (E l , e l ) becomes the new input set of the deep learning model B2. The input set is divided for model training. The Markov chain-Monte Carlo sampling method is used to screen the optimal candidate set in each iteration. The weights and thresholds are adjusted after training, and finally the optimized drilling and extraction process parameters are obtained, which are verified using the validation set.

[0026] Furthermore, the geological parameters described in step S1 include coal reservoir physical parameters: coal seam temperature, coal body density, elastic modulus of coal matrix, coal seam gas content, ground stress, initial fracture permeability, initial matrix porosity, initial fracture porosity, adsorption volume and adsorption pressure, and initial gas pressure; the engineering parameters include fracturing design parameters, gas extraction parameters, and the input data set U i are all fracturing design parameters, input data set U k The parameters in the table are gas extraction parameters.

[0027] Furthermore, the gas extraction parameters include hole diameter, hole depth, hole spacing, inclination angle, pre-extraction time, gas extraction quantity and gas extraction concentration.

[0028] Furthermore, the fracturing parameters to be inverted in step S1 include important factors affecting the fracturing effect, such as the number of effective perforations and the drag reduction coefficient.

[0029] Furthermore, the gas extraction process parameters to be inverted in step S1 include extraction negative pressure, borehole diameter, hole spacing, inclination, hole depth and shortest pre-extraction time, which are important factors affecting the gas extraction effect of the coal seam.

[0030] Furthermore, the assumptions in step S1 include the following aspects: coal seam isotropy / anisotropy, homogeneity / inhomogeneity, elastoplastic assumption, deformation assumption, isothermal / non-isothermal, gas composition, gas is in an ideal / non-ideal state, and gas is in a saturated / unsaturated state.

[0031] Furthermore, the fracture propagation model in step S1 includes natural fractures and hydraulic fractures, the natural fractures are generated by a discrete fracture network model, and the hydraulic fractures are simulated by a DDM method.

[0032] Furthermore, in the EDFM model in step S5, the matrix grid and the crack grid are connected by non-adjacent connection pairs (NNCs), the matrix adopts a structured grid, and the crack is simulated as a rectangular plate embedded in the matrix grid.

[0033] Furthermore, the process of hard constraint mapping of domain knowledge described in steps S2 and S6 is as follows: randomly obtain target points in the input data set, determine the range of the local constraint block, determine the physical control equation according to the domain knowledge, discretize the partial differential equation and perform matrix decomposition in the constraint block, obtain the constraint matrix and the projection matrix, and convert the data set generated by the simulation into (i.e., the prediction matrix) is projected into the constraint block to obtain the prediction result that satisfies the physical constraints.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] First, physical laws such as fracture extension law, desorption and adsorption process law, diffusion equation, seepage law, slippage effect and physical information such as coal reservoir reservoir characteristics are integrated into the coal seam geological model constructed by numerical simulation software, and combined with the advantages of deep learning methods in effectively extracting data laws, the limited historical experience cases and simulation cases are fully studied, which can extract effective information in the gas extraction of this coal seam, and can ensure that the learned results are consistent with both the laws of physical knowledge and the actual distribution, thereby improving the interpretability of the obtained results;

[0036] Second, the historical gas concentration data is used to optimize the fracturing parameters in the initial numerical model, and the fracture morphology of the geological model is corrected by back-substituting it into the numerical model. The gas extraction process parameters and simulated gas scalar data sets that conform to the laws of physics are generated with the help of the corrected geological model. Finally, the gas extraction scalar is used to optimize the gas extraction process parameters. Compared with the single optimization method, the double-layer optimization method incorporates physical mechanisms, so that the optimization results are consistent with both data laws and physical reality.

[0037] 3. Since the numerical simulation process will inevitably introduce prediction errors, and the hard constraint mapping (HCP) method gives the nearest solution that conforms to the original equation on the local constraint block from a mathematical perspective, it is considered to use domain knowledge to perform hard constraint mapping on the simulation results to reduce the deviation caused by the physical simulation process, so that the model can satisfy the physical laws on the local constraint block, which can improve the prediction accuracy of the overall model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the present invention;

[0039] Figure 2 Flow chart of crack morphology of the model for gas concentration correction;

[0040] Figure 3 Flowchart of extraction process parameters for optimizing gas extraction volume. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0042] like Figures 1 to 3 The double-layer optimization method of drilling and extraction parameters based on numerical simulation and deep learning shown in the figure includes the following steps:

[0043] S1. Establish initial numerical model A1:

[0044] Collect geological parameters and engineering parameter data. The geological parameters include coal reservoir physical parameters: coal seam temperature, coal body density, elastic modulus of coal matrix, coal seam gas content, ground stress, initial fracture permeability, initial matrix porosity, initial fracture porosity, adsorption volume and adsorption pressure, and initial gas pressure; the engineering parameters include fracturing design parameters, gas extraction parameters (including hole diameter, hole depth, hole spacing, inclination, pre-extraction time, gas extraction pure amount, and gas extraction concentration);

[0045] Determine the fracturing parameters to be inverted (including the number of effective perforations and the drag reduction coefficient), the gas extraction process parameters to be inverted (including the extraction negative pressure, borehole diameter, hole spacing, inclination, hole depth and the shortest pre-extraction time) and their empirical ranges, and generate the input data set U through LHS sampling. i and U k , input data set U i are all fracturing design parameters, input data set U k The middle ones are gas extraction parameters;

[0046] Set the geometric model parameters, boundary conditions, initial conditions, and assumptions (including the following aspects: isotropy / anisotropy of coal seams, homogeneity / inhomogeneity, elastic-plastic assumption, deformation assumption, isothermal / non-isothermal, gas composition (considering only gas / gas-air mixture), gas in ideal / non-ideal state, gas in saturated / unsaturated state), determine the mathematical model of gas drilling extraction based on gas desorption and adsorption characteristics, molecular diffusion, seepage characteristics, and slippage effect, import various parameters, conditions, and geometric models into COMSOL or FLUENT numerical simulation software, import the fracture propagation model (including natural fractures and hydraulic fractures, natural fractures are generated by discrete fracture network model, and hydraulic fractures are simulated by DDM method), and establish the initial numerical simulation model A1;

[0047] S2, the simulation sample set I1 and the experience sample set I2 that form the deep learning model:

[0048] After numerical simulation, the gas concentration distribution data set c is generated i * , for the dataset c i * Perform hard constraint mapping of domain knowledge and generate a gas concentration dataset c that complies with physical law constraints in the local constraint block. i , input data set U i and gas concentration dataset c i Combined into simulation sample set I1: (U i , c i ); Based on past historical experience, we form an experience sample set I2: (E j , e j ), E j is the range of geological parameters and fracturing parameters in historical data, e j is the gas concentration distribution in historical data;

[0049] S3. Build a deep learning model B1 based on the simulated sample set I1 and the empirical sample set I2, and improve the loss function:

[0050] A fully connected neural network is used, and the initial number of hidden layers, number of units, optimization algorithm, learning rate and activation function are set. The hyperparameters are continuously optimized iteratively during training, and the loss function L is:

[0051] L=λL M +μL E ; (1)

[0052] In formula (1), L M is the simulated sample loss, λ is the simulated sample loss weight coefficient; L E is the experience sample loss, μ is the experience sample loss weight coefficient; λ and μ are both constants, which are experimentally adjusted during the training of the deep learning model B1;

[0053] Simulated sample loss L M According to the partial differential equation in the hard constraint mapping, it is obtained as shown in formula (2):

[0054]

[0055] In formula (2), is the deviation of the ith partial differential equation, are all the variables in the ith partial differential equation system, N f is the number of partial differential equations;

[0056] Experience sample loss L E Calculate according to formula (3):

[0057]

[0058] In formula (3), e j , are the observed value and predicted value of the empirical sample set, respectively, and n is the number of empirical samples;

[0059] S4, model training and verification: The simulation sample set I1 and the experience sample set I2 are combined into a training set and a verification set, and the training set is input into the deep learning model B1 for training to invert the input parameters. The MCMC method is used to screen the optimal candidate set in each iteration. After the training is completed, the verification set is used for verification, and finally the optimal fracturing parameter combination is obtained by inversion;

[0060] S5, correcting the fracture morphology: inputting the optimal fracturing parameter combination obtained in step S4 into the initial numerical model A1 established in step S1, generating a numerical model A2 after correcting the fracture morphology, and connecting the numerical model A2 with the EDFM model to form a fracture network; the EDFM model connects the matrix grid and the fracture grid through NNC, the matrix adopts a structured grid, and the fracture is simulated as a rectangular plate embedded in the matrix grid;

[0061] S6. Construct simulated sample set I3: Input data set U k Input the numerical simulation software and use the numerical model A2 established in step S5 to simulate and generate the pure amount of gas extraction. The same method as step S2 is used to extract the pure amount of gas. Perform hard constraint mapping of domain knowledge to generate a gas extraction scalar distribution q that meets physical constraints k , combined input dataset U k And the simulation results of gas extraction scalar distribution q k For the simulated data set I3: (U k ,q k );

[0062] S7, build a deep learning model B2, and obtain the optimal gas extraction process parameters after inversion: Use the same method as step S3 to build a deep learning model B2 and simulate the data set I3: (U k ,q k )Combined with empirical data set I4: (E l , e l ) becomes the new input set of the deep learning model B2. The input set is divided for model training. The Markov chain-Monte Carlo sampling method is used to screen the optimal candidate set in each iteration. The weights and thresholds are adjusted after training, and finally the optimized drilling and extraction process parameters are obtained, which are verified using the validation set.

[0063] Furthermore, the process of hard constraint mapping of domain knowledge described in steps S2 and S6 is as follows: randomly obtain target points in the input data set, determine the range of the local constraint block, determine the physical control equation according to the domain knowledge, discretize the partial differential equation and perform matrix decomposition in the constraint block, obtain the constraint matrix and the projection matrix, and convert the data set generated by the simulation into (i.e., the prediction matrix) is projected into the constraint block to obtain the prediction result that satisfies the physical constraints.

[0064] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A two-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning, characterized in that: The following steps are involved: S1. Establish initial numerical model A1: Collect geological parameters and engineering parameter data, determine the fracturing parameters to be inverted, the gas extraction process parameters to be inverted and their empirical ranges, and generate the input data set U through LHS sampling i and U k ; Set geometric model parameters, boundary conditions, initial conditions, and assumptions, determine the mathematical model of gas drilling extraction based on gas desorption and adsorption characteristics, molecular diffusion, seepage characteristics, and slippage effects, import various parameters, conditions, and geometric models into COMSOL or FLUENT numerical simulation software, import the fracture propagation model, and establish the initial numerical simulation model A1; S2, the simulation sample set I1 and the experience sample set I2 that form the deep learning model: After numerical simulation, the gas concentration distribution data set c is generated i * , for the dataset c i * Perform hard constraint mapping of domain knowledge and generate a gas concentration dataset c that complies with physical law constraints in the local constraint block. i , input data set U i and gas concentration dataset c i Combined into simulation sample set I1: (U i , c i ); Based on past historical experience, we form an experience sample set I2: (E j , e j ), E j is the range of geological parameters and fracturing parameters in historical data, e j is the gas concentration distribution in historical data; S3. Build a deep learning model B1 based on the simulated sample set I1 and the empirical sample set I2, and improve the loss function: A fully connected neural network is used, and the initial number of hidden layers, number of units, optimization algorithm, learning rate and activation function are set. The hyperparameters are continuously optimized iteratively during training, and the loss function L is: L=λL M +μL E ; (1) In formula (1), L M is the simulated sample loss, λ is the simulated sample loss weight coefficient; L E is the experience sample loss, μ is the experience sample loss weight coefficient; λ and μ are both constants, which are experimentally adjusted during the training of the deep learning model B1; Simulated sample loss L M According to the partial differential equation in the hard constraint mapping, it is obtained as shown in formula (2): In formula (2), is the deviation of the ith partial differential equation, are all the variables in the ith partial differential equation system, N f is the number of partial differential equations; Experience sample loss L E Calculate according to formula (3): In formula (3), e j , are the observed value and predicted value of the empirical sample set, respectively, and n is the number of empirical samples; S4, model training and verification: The simulation sample set I1 and the experience sample set I2 are combined into a training set and a verification set, and the training set is input into the deep learning model B1 for training to invert the input parameters. The Markov chain-Monte Carlo sampling method is used to screen the optimal candidate set in each iteration. After the training is completed, the verification set is used for verification, and finally the optimal fracturing parameter combination is inverted; S5, correcting the fracture morphology: inputting the optimal fracturing parameter combination obtained in step S4 into the initial numerical model A1 established in step S1, generating a numerical model A2 after correcting the fracture morphology, and connecting the numerical model A2 with the EDFM model to form a fracture network; S6. Construct simulated sample set I3: Input data set U k Input the numerical simulation software and use the numerical model A2 established in step S5 to simulate and generate the pure amount of gas extraction. The same method as step S2 is used to extract the pure amount of gas. Perform hard constraint mapping of domain knowledge to generate a gas extraction scalar distribution q that meets physical constraints k , combined input dataset U k And the simulation results of gas extraction scalar distribution q k For the simulated data set I3: (U k ,q k ); S7, build a deep learning model B2, and obtain the optimal gas extraction process parameters after inversion: Use the same method as step S3 to build a deep learning model B2 and simulate the data set I3: (U k ,q k )Combined with empirical data set I4: (E l , e l ) becomes the new input set of the deep learning model B2. The input set is divided for model training. The Markov chain-Monte Carlo sampling method is used to screen the optimal candidate set in each iteration. The weights and thresholds are adjusted after training, and finally the optimized drilling and extraction process parameters are obtained, which are verified using the validation set.

2. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 1 is characterized in that: The geological parameters described in step S1 include coal reservoir physical parameters: coal seam temperature, coal body density, elastic modulus of coal matrix, coal seam gas content, ground stress, initial fracture permeability, initial matrix porosity, initial fracture porosity, adsorption volume and adsorption pressure, and initial gas pressure; the engineering parameters include fracturing design parameters and gas extraction parameters, and the input data set U i are all fracturing design parameters, input data set U k The parameters in the table are gas extraction parameters.

3. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 2 is characterized in that: The gas extraction parameters include hole diameter, hole depth, hole spacing, inclination angle, pre-extraction time, gas extraction quantity and gas extraction concentration.

4. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 1 is characterized in that: The fracturing parameters to be inverted in step S1 include the number of effective perforations and the drag reduction coefficient.

5. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 1 is characterized in that: The gas extraction process parameters to be inverted in step S1 include extraction negative pressure, borehole diameter, hole spacing, inclination angle, hole depth and shortest pre-extraction time.

6. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 1 is characterized in that: The assumptions in step S1 include the following aspects: coal seam isotropy / anisotropy, homogeneity / inhomogeneity, elastic-plastic assumption, deformation assumption, isothermal / non-isothermal, gas composition, gas is in an ideal / non-ideal state, and gas is in a saturated / unsaturated state.

7. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 1 is characterized in that: The fracture extension model in step S1 includes natural fractures and hydraulic fractures. Natural fractures are generated by discrete fracture network model, and hydraulic fractures are simulated by DDM method.

8. The method for optimizing drilling and extraction parameters based on numerical simulation and deep learning according to claim 1, characterized in that: In step S5, the EDFM model connects the matrix grid and the crack grid through non-adjacent connection pairs, the matrix adopts a structured grid, and the crack is simulated as a rectangular plate embedded in the matrix grid.

9. The double-layer optimization method for drilling and extraction parameters based on numerical simulation and deep learning according to claim 1, characterized in that: The process of hard constraint mapping of domain knowledge described in steps S2 and S6 is as follows: randomly obtain target points in the input data set, determine the range of the local constraint block, determine the physical control equation according to the domain knowledge, discretize the partial differential equation and perform matrix decomposition in the constraint block, obtain the constraint matrix and the projection matrix, and convert the data set generated by the simulation into Project it into the constraint block to obtain the prediction results that meet the physical constraints.

Images